| 1 | //===---- CGOpenMPRuntimeGPU.cpp - Interface to OpenMP GPU Runtimes ----===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This provides a generalized class for OpenMP runtime code generation |
| 10 | // specialized by GPU targets NVPTX, AMDGCN and SPIR-V. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "CGOpenMPRuntimeGPU.h" |
| 15 | #include "CGDebugInfo.h" |
| 16 | #include "CodeGenFunction.h" |
| 17 | #include "TargetInfo.h" |
| 18 | #include "clang/AST/Attr.h" |
| 19 | #include "clang/AST/DeclOpenMP.h" |
| 20 | #include "clang/AST/OpenMPClause.h" |
| 21 | #include "clang/AST/StmtOpenMP.h" |
| 22 | #include "clang/AST/StmtVisitor.h" |
| 23 | #include "llvm/ADT/SmallPtrSet.h" |
| 24 | #include "llvm/Frontend/OpenMP/OMPDeviceConstants.h" |
| 25 | #include "llvm/Frontend/OpenMP/OMPGridValues.h" |
| 26 | #include "llvm/IR/IRBuilder.h" |
| 27 | #include "llvm/IR/Instructions.h" |
| 28 | #include "llvm/TargetParser/NVPTXTargetParser.h" |
| 29 | |
| 30 | using namespace clang; |
| 31 | using namespace CodeGen; |
| 32 | using namespace llvm::omp; |
| 33 | |
| 34 | namespace { |
| 35 | /// Pre(post)-action for different OpenMP constructs specialized for NVPTX. |
| 36 | class NVPTXActionTy final : public PrePostActionTy { |
| 37 | llvm::FunctionCallee EnterCallee = nullptr; |
| 38 | ArrayRef<llvm::Value *> EnterArgs; |
| 39 | llvm::FunctionCallee ExitCallee = nullptr; |
| 40 | ArrayRef<llvm::Value *> ExitArgs; |
| 41 | bool Conditional = false; |
| 42 | llvm::BasicBlock *ContBlock = nullptr; |
| 43 | |
| 44 | public: |
| 45 | NVPTXActionTy(llvm::FunctionCallee EnterCallee, |
| 46 | ArrayRef<llvm::Value *> EnterArgs, |
| 47 | llvm::FunctionCallee ExitCallee, |
| 48 | ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) |
| 49 | : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), |
| 50 | ExitArgs(ExitArgs), Conditional(Conditional) {} |
| 51 | void Enter(CodeGenFunction &CGF) override { |
| 52 | llvm::Value *EnterRes = CGF.EmitRuntimeCall(callee: EnterCallee, args: EnterArgs); |
| 53 | if (Conditional) { |
| 54 | llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(Arg: EnterRes); |
| 55 | auto *ThenBlock = CGF.createBasicBlock(name: "omp_if.then" ); |
| 56 | ContBlock = CGF.createBasicBlock(name: "omp_if.end" ); |
| 57 | // Generate the branch (If-stmt) |
| 58 | CGF.Builder.CreateCondBr(Cond: CallBool, True: ThenBlock, False: ContBlock); |
| 59 | CGF.EmitBlock(BB: ThenBlock); |
| 60 | } |
| 61 | } |
| 62 | void Done(CodeGenFunction &CGF) { |
| 63 | // Emit the rest of blocks/branches |
| 64 | CGF.EmitBranch(Block: ContBlock); |
| 65 | CGF.EmitBlock(BB: ContBlock, IsFinished: true); |
| 66 | } |
| 67 | void Exit(CodeGenFunction &CGF) override { |
| 68 | CGF.EmitRuntimeCall(callee: ExitCallee, args: ExitArgs); |
| 69 | } |
| 70 | }; |
| 71 | |
| 72 | /// A class to track the execution mode when codegening directives within |
| 73 | /// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry |
| 74 | /// to the target region and used by containing directives such as 'parallel' |
| 75 | /// to emit optimized code. |
| 76 | class ExecutionRuntimeModesRAII { |
| 77 | private: |
| 78 | CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode = |
| 79 | CGOpenMPRuntimeGPU::EM_Unknown; |
| 80 | CGOpenMPRuntimeGPU::ExecutionMode &ExecMode; |
| 81 | |
| 82 | public: |
| 83 | ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode, |
| 84 | CGOpenMPRuntimeGPU::ExecutionMode EntryMode) |
| 85 | : ExecMode(ExecMode) { |
| 86 | SavedExecMode = ExecMode; |
| 87 | ExecMode = EntryMode; |
| 88 | } |
| 89 | ~ExecutionRuntimeModesRAII() { ExecMode = SavedExecMode; } |
| 90 | }; |
| 91 | |
| 92 | static const ValueDecl *getPrivateItem(const Expr *RefExpr) { |
| 93 | RefExpr = RefExpr->IgnoreParens(); |
| 94 | if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(Val: RefExpr)) { |
| 95 | const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); |
| 96 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Val: Base)) |
| 97 | Base = TempASE->getBase()->IgnoreParenImpCasts(); |
| 98 | RefExpr = Base; |
| 99 | } else if (auto *OASE = dyn_cast<ArraySectionExpr>(Val: RefExpr)) { |
| 100 | const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); |
| 101 | while (const auto *TempOASE = dyn_cast<ArraySectionExpr>(Val: Base)) |
| 102 | Base = TempOASE->getBase()->IgnoreParenImpCasts(); |
| 103 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Val: Base)) |
| 104 | Base = TempASE->getBase()->IgnoreParenImpCasts(); |
| 105 | RefExpr = Base; |
| 106 | } |
| 107 | RefExpr = RefExpr->IgnoreParenImpCasts(); |
| 108 | if (const auto *DE = dyn_cast<DeclRefExpr>(Val: RefExpr)) |
| 109 | return cast<ValueDecl>(Val: DE->getDecl()->getCanonicalDecl()); |
| 110 | const auto *ME = cast<MemberExpr>(Val: RefExpr); |
| 111 | return cast<ValueDecl>(Val: ME->getMemberDecl()->getCanonicalDecl()); |
| 112 | } |
| 113 | |
| 114 | static RecordDecl *buildRecordForGlobalizedVars( |
| 115 | ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls, |
| 116 | ArrayRef<const ValueDecl *> EscapedDeclsForTeams, |
| 117 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 118 | &MappedDeclsFields, |
| 119 | int BufSize) { |
| 120 | using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>; |
| 121 | if (EscapedDecls.empty() && EscapedDeclsForTeams.empty()) |
| 122 | return nullptr; |
| 123 | SmallVector<VarsDataTy, 4> GlobalizedVars; |
| 124 | for (const ValueDecl *D : EscapedDecls) |
| 125 | GlobalizedVars.emplace_back(Args: C.getDeclAlign(D), Args&: D); |
| 126 | for (const ValueDecl *D : EscapedDeclsForTeams) |
| 127 | GlobalizedVars.emplace_back(Args: C.getDeclAlign(D), Args&: D); |
| 128 | |
| 129 | // Build struct _globalized_locals_ty { |
| 130 | // /* globalized vars */[WarSize] align (decl_align) |
| 131 | // /* globalized vars */ for EscapedDeclsForTeams |
| 132 | // }; |
| 133 | RecordDecl *GlobalizedRD = C.buildImplicitRecord(Name: "_globalized_locals_ty" ); |
| 134 | GlobalizedRD->startDefinition(); |
| 135 | llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped(llvm::from_range, |
| 136 | EscapedDeclsForTeams); |
| 137 | for (const auto &Pair : GlobalizedVars) { |
| 138 | const ValueDecl *VD = Pair.second; |
| 139 | QualType Type = VD->getType(); |
| 140 | if (Type->isLValueReferenceType()) |
| 141 | Type = C.getPointerType(T: Type.getNonReferenceType()); |
| 142 | else |
| 143 | Type = Type.getNonReferenceType(); |
| 144 | SourceLocation Loc = VD->getLocation(); |
| 145 | FieldDecl *Field; |
| 146 | if (SingleEscaped.count(Ptr: VD)) { |
| 147 | Field = FieldDecl::Create( |
| 148 | C, DC: GlobalizedRD, StartLoc: Loc, IdLoc: Loc, Id: VD->getIdentifier(), T: Type, |
| 149 | TInfo: C.getTrivialTypeSourceInfo(T: Type, Loc: SourceLocation()), |
| 150 | /*BW=*/nullptr, /*Mutable=*/false, |
| 151 | /*InitStyle=*/ICIS_NoInit); |
| 152 | Field->setAccess(AS_public); |
| 153 | if (VD->hasAttrs()) { |
| 154 | for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), |
| 155 | E(VD->getAttrs().end()); |
| 156 | I != E; ++I) |
| 157 | Field->addAttr(A: *I); |
| 158 | } |
| 159 | } else { |
| 160 | if (BufSize > 1) { |
| 161 | llvm::APInt ArraySize(32, BufSize); |
| 162 | Type = C.getConstantArrayType(EltTy: Type, ArySize: ArraySize, SizeExpr: nullptr, |
| 163 | ASM: ArraySizeModifier::Normal, IndexTypeQuals: 0); |
| 164 | } |
| 165 | Field = FieldDecl::Create( |
| 166 | C, DC: GlobalizedRD, StartLoc: Loc, IdLoc: Loc, Id: VD->getIdentifier(), T: Type, |
| 167 | TInfo: C.getTrivialTypeSourceInfo(T: Type, Loc: SourceLocation()), |
| 168 | /*BW=*/nullptr, /*Mutable=*/false, |
| 169 | /*InitStyle=*/ICIS_NoInit); |
| 170 | Field->setAccess(AS_public); |
| 171 | llvm::APInt Align(32, Pair.first.getQuantity()); |
| 172 | Field->addAttr(A: AlignedAttr::CreateImplicit( |
| 173 | Ctx&: C, /*IsAlignmentExpr=*/true, |
| 174 | Alignment: IntegerLiteral::Create(C, V: Align, |
| 175 | type: C.getIntTypeForBitwidth(DestWidth: 32, /*Signed=*/0), |
| 176 | l: SourceLocation()), |
| 177 | Range: {}, S: AlignedAttr::GNU_aligned)); |
| 178 | } |
| 179 | GlobalizedRD->addDecl(D: Field); |
| 180 | MappedDeclsFields.try_emplace(Key: VD, Args&: Field); |
| 181 | } |
| 182 | GlobalizedRD->completeDefinition(); |
| 183 | return GlobalizedRD; |
| 184 | } |
| 185 | |
| 186 | /// Get the list of variables that can escape their declaration context. |
| 187 | class CheckVarsEscapingDeclContext final |
| 188 | : public ConstStmtVisitor<CheckVarsEscapingDeclContext> { |
| 189 | CodeGenFunction &CGF; |
| 190 | llvm::SetVector<const ValueDecl *> EscapedDecls; |
| 191 | llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls; |
| 192 | llvm::SetVector<const ValueDecl *> DelayedVariableLengthDecls; |
| 193 | llvm::SmallPtrSet<const Decl *, 4> EscapedParameters; |
| 194 | RecordDecl *GlobalizedRD = nullptr; |
| 195 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; |
| 196 | bool AllEscaped = false; |
| 197 | bool IsForCombinedParallelRegion = false; |
| 198 | |
| 199 | void markAsEscaped(const ValueDecl *VD) { |
| 200 | // Do not globalize declare target variables. |
| 201 | if (!isa<VarDecl>(Val: VD) || |
| 202 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) |
| 203 | return; |
| 204 | VD = cast<ValueDecl>(Val: VD->getCanonicalDecl()); |
| 205 | // Use user-specified allocation. |
| 206 | if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>()) |
| 207 | return; |
| 208 | // Variables captured by value must be globalized. |
| 209 | bool IsCaptured = false; |
| 210 | if (auto *CSI = CGF.CapturedStmtInfo) { |
| 211 | if (const FieldDecl *FD = CSI->lookup(VD: cast<VarDecl>(Val: VD))) { |
| 212 | // Check if need to capture the variable that was already captured by |
| 213 | // value in the outer region. |
| 214 | IsCaptured = true; |
| 215 | if (!IsForCombinedParallelRegion) { |
| 216 | if (!FD->hasAttrs()) |
| 217 | return; |
| 218 | const auto *Attr = FD->getAttr<OMPCaptureKindAttr>(); |
| 219 | if (!Attr) |
| 220 | return; |
| 221 | if (((Attr->getCaptureKind() != OMPC_map) && |
| 222 | !isOpenMPPrivate(Kind: Attr->getCaptureKind())) || |
| 223 | ((Attr->getCaptureKind() == OMPC_map) && |
| 224 | !FD->getType()->isAnyPointerType())) |
| 225 | return; |
| 226 | } |
| 227 | if (!FD->getType()->isReferenceType()) { |
| 228 | assert(!VD->getType()->isVariablyModifiedType() && |
| 229 | "Parameter captured by value with variably modified type" ); |
| 230 | EscapedParameters.insert(Ptr: VD); |
| 231 | } else if (!IsForCombinedParallelRegion) { |
| 232 | return; |
| 233 | } |
| 234 | } |
| 235 | } |
| 236 | if ((!CGF.CapturedStmtInfo || |
| 237 | (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) && |
| 238 | VD->getType()->isReferenceType()) |
| 239 | // Do not globalize variables with reference type. |
| 240 | return; |
| 241 | if (VD->getType()->isVariablyModifiedType()) { |
| 242 | // If not captured at the target region level then mark the escaped |
| 243 | // variable as delayed. |
| 244 | if (IsCaptured) |
| 245 | EscapedVariableLengthDecls.insert(X: VD); |
| 246 | else |
| 247 | DelayedVariableLengthDecls.insert(X: VD); |
| 248 | } else |
| 249 | EscapedDecls.insert(X: VD); |
| 250 | } |
| 251 | |
| 252 | void VisitValueDecl(const ValueDecl *VD) { |
| 253 | if (VD->getType()->isLValueReferenceType()) |
| 254 | markAsEscaped(VD); |
| 255 | if (const auto *VarD = dyn_cast<VarDecl>(Val: VD)) { |
| 256 | if (!isa<ParmVarDecl>(Val: VarD) && VarD->hasInit()) { |
| 257 | const bool SavedAllEscaped = AllEscaped; |
| 258 | AllEscaped = VD->getType()->isLValueReferenceType(); |
| 259 | Visit(S: VarD->getInit()); |
| 260 | AllEscaped = SavedAllEscaped; |
| 261 | } |
| 262 | } |
| 263 | } |
| 264 | void VisitOpenMPCapturedStmt(const CapturedStmt *S, |
| 265 | ArrayRef<OMPClause *> Clauses, |
| 266 | bool IsCombinedParallelRegion) { |
| 267 | if (!S) |
| 268 | return; |
| 269 | for (const CapturedStmt::Capture &C : S->captures()) { |
| 270 | if (C.capturesVariable() && !C.capturesVariableByCopy()) { |
| 271 | const ValueDecl *VD = C.getCapturedVar(); |
| 272 | bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion; |
| 273 | if (IsCombinedParallelRegion) { |
| 274 | // Check if the variable is privatized in the combined construct and |
| 275 | // those private copies must be shared in the inner parallel |
| 276 | // directive. |
| 277 | IsForCombinedParallelRegion = false; |
| 278 | for (const OMPClause *C : Clauses) { |
| 279 | if (!isOpenMPPrivate(Kind: C->getClauseKind()) || |
| 280 | C->getClauseKind() == OMPC_reduction || |
| 281 | C->getClauseKind() == OMPC_linear || |
| 282 | C->getClauseKind() == OMPC_private) |
| 283 | continue; |
| 284 | ArrayRef<const Expr *> Vars; |
| 285 | if (const auto *PC = dyn_cast<OMPFirstprivateClause>(Val: C)) |
| 286 | Vars = PC->getVarRefs(); |
| 287 | else if (const auto *PC = dyn_cast<OMPLastprivateClause>(Val: C)) |
| 288 | Vars = PC->getVarRefs(); |
| 289 | else |
| 290 | llvm_unreachable("Unexpected clause." ); |
| 291 | for (const auto *E : Vars) { |
| 292 | const Decl *D = |
| 293 | cast<DeclRefExpr>(Val: E)->getDecl()->getCanonicalDecl(); |
| 294 | if (D == VD->getCanonicalDecl()) { |
| 295 | IsForCombinedParallelRegion = true; |
| 296 | break; |
| 297 | } |
| 298 | } |
| 299 | if (IsForCombinedParallelRegion) |
| 300 | break; |
| 301 | } |
| 302 | } |
| 303 | markAsEscaped(VD); |
| 304 | if (isa<OMPCapturedExprDecl>(Val: VD)) |
| 305 | VisitValueDecl(VD); |
| 306 | IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion; |
| 307 | } |
| 308 | } |
| 309 | } |
| 310 | |
| 311 | void buildRecordForGlobalizedVars(bool IsInTTDRegion) { |
| 312 | assert(!GlobalizedRD && |
| 313 | "Record for globalized variables is built already." ); |
| 314 | ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams; |
| 315 | unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; |
| 316 | if (IsInTTDRegion) |
| 317 | EscapedDeclsForTeams = EscapedDecls.getArrayRef(); |
| 318 | else |
| 319 | EscapedDeclsForParallel = EscapedDecls.getArrayRef(); |
| 320 | GlobalizedRD = ::buildRecordForGlobalizedVars( |
| 321 | C&: CGF.getContext(), EscapedDecls: EscapedDeclsForParallel, EscapedDeclsForTeams, |
| 322 | MappedDeclsFields, BufSize: WarpSize); |
| 323 | } |
| 324 | |
| 325 | public: |
| 326 | CheckVarsEscapingDeclContext(CodeGenFunction &CGF, |
| 327 | ArrayRef<const ValueDecl *> TeamsReductions) |
| 328 | : CGF(CGF), EscapedDecls(llvm::from_range, TeamsReductions) {} |
| 329 | ~CheckVarsEscapingDeclContext() = default; |
| 330 | void VisitDeclStmt(const DeclStmt *S) { |
| 331 | if (!S) |
| 332 | return; |
| 333 | for (const Decl *D : S->decls()) |
| 334 | if (const auto *VD = dyn_cast_or_null<ValueDecl>(Val: D)) |
| 335 | VisitValueDecl(VD); |
| 336 | } |
| 337 | void VisitOMPExecutableDirective(const OMPExecutableDirective *D) { |
| 338 | if (!D) |
| 339 | return; |
| 340 | if (!D->hasAssociatedStmt()) |
| 341 | return; |
| 342 | if (const auto *S = |
| 343 | dyn_cast_or_null<CapturedStmt>(Val: D->getAssociatedStmt())) { |
| 344 | // Do not analyze directives that do not actually require capturing, |
| 345 | // like `omp for` or `omp simd` directives. |
| 346 | llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; |
| 347 | getOpenMPCaptureRegions(CaptureRegions, DKind: D->getDirectiveKind()); |
| 348 | if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) { |
| 349 | VisitStmt(S: S->getCapturedStmt()); |
| 350 | return; |
| 351 | } |
| 352 | VisitOpenMPCapturedStmt( |
| 353 | S, Clauses: D->clauses(), |
| 354 | IsCombinedParallelRegion: CaptureRegions.back() == OMPD_parallel && |
| 355 | isOpenMPDistributeDirective(DKind: D->getDirectiveKind())); |
| 356 | } |
| 357 | } |
| 358 | void VisitCapturedStmt(const CapturedStmt *S) { |
| 359 | if (!S) |
| 360 | return; |
| 361 | for (const CapturedStmt::Capture &C : S->captures()) { |
| 362 | if (C.capturesVariable() && !C.capturesVariableByCopy()) { |
| 363 | const ValueDecl *VD = C.getCapturedVar(); |
| 364 | markAsEscaped(VD); |
| 365 | if (isa<OMPCapturedExprDecl>(Val: VD)) |
| 366 | VisitValueDecl(VD); |
| 367 | } |
| 368 | } |
| 369 | } |
| 370 | void VisitLambdaExpr(const LambdaExpr *E) { |
| 371 | if (!E) |
| 372 | return; |
| 373 | for (const LambdaCapture &C : E->captures()) { |
| 374 | if (C.capturesVariable()) { |
| 375 | if (C.getCaptureKind() == LCK_ByRef) { |
| 376 | const ValueDecl *VD = C.getCapturedVar(); |
| 377 | markAsEscaped(VD); |
| 378 | if (E->isInitCapture(Capture: &C) || isa<OMPCapturedExprDecl>(Val: VD)) |
| 379 | VisitValueDecl(VD); |
| 380 | } |
| 381 | } |
| 382 | } |
| 383 | } |
| 384 | void VisitBlockExpr(const BlockExpr *E) { |
| 385 | if (!E) |
| 386 | return; |
| 387 | for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) { |
| 388 | if (C.isByRef()) { |
| 389 | const VarDecl *VD = C.getVariable(); |
| 390 | markAsEscaped(VD); |
| 391 | if (isa<OMPCapturedExprDecl>(Val: VD) || VD->isInitCapture()) |
| 392 | VisitValueDecl(VD); |
| 393 | } |
| 394 | } |
| 395 | } |
| 396 | void VisitCallExpr(const CallExpr *E) { |
| 397 | if (!E) |
| 398 | return; |
| 399 | for (const Expr *Arg : E->arguments()) { |
| 400 | if (!Arg) |
| 401 | continue; |
| 402 | if (Arg->isLValue()) { |
| 403 | const bool SavedAllEscaped = AllEscaped; |
| 404 | AllEscaped = true; |
| 405 | Visit(S: Arg); |
| 406 | AllEscaped = SavedAllEscaped; |
| 407 | } else { |
| 408 | Visit(S: Arg); |
| 409 | } |
| 410 | } |
| 411 | Visit(S: E->getCallee()); |
| 412 | } |
| 413 | void VisitDeclRefExpr(const DeclRefExpr *E) { |
| 414 | if (!E) |
| 415 | return; |
| 416 | const ValueDecl *VD = E->getDecl(); |
| 417 | if (AllEscaped) |
| 418 | markAsEscaped(VD); |
| 419 | if (isa<OMPCapturedExprDecl>(Val: VD)) |
| 420 | VisitValueDecl(VD); |
| 421 | else if (VD->isInitCapture()) |
| 422 | VisitValueDecl(VD); |
| 423 | } |
| 424 | void VisitUnaryOperator(const UnaryOperator *E) { |
| 425 | if (!E) |
| 426 | return; |
| 427 | if (E->getOpcode() == UO_AddrOf) { |
| 428 | const bool SavedAllEscaped = AllEscaped; |
| 429 | AllEscaped = true; |
| 430 | Visit(S: E->getSubExpr()); |
| 431 | AllEscaped = SavedAllEscaped; |
| 432 | } else { |
| 433 | Visit(S: E->getSubExpr()); |
| 434 | } |
| 435 | } |
| 436 | void VisitImplicitCastExpr(const ImplicitCastExpr *E) { |
| 437 | if (!E) |
| 438 | return; |
| 439 | if (E->getCastKind() == CK_ArrayToPointerDecay) { |
| 440 | const bool SavedAllEscaped = AllEscaped; |
| 441 | AllEscaped = true; |
| 442 | Visit(S: E->getSubExpr()); |
| 443 | AllEscaped = SavedAllEscaped; |
| 444 | } else { |
| 445 | Visit(S: E->getSubExpr()); |
| 446 | } |
| 447 | } |
| 448 | void VisitExpr(const Expr *E) { |
| 449 | if (!E) |
| 450 | return; |
| 451 | bool SavedAllEscaped = AllEscaped; |
| 452 | if (!E->isLValue()) |
| 453 | AllEscaped = false; |
| 454 | for (const Stmt *Child : E->children()) |
| 455 | if (Child) |
| 456 | Visit(S: Child); |
| 457 | AllEscaped = SavedAllEscaped; |
| 458 | } |
| 459 | void VisitStmt(const Stmt *S) { |
| 460 | if (!S) |
| 461 | return; |
| 462 | for (const Stmt *Child : S->children()) |
| 463 | if (Child) |
| 464 | Visit(S: Child); |
| 465 | } |
| 466 | |
| 467 | /// Returns the record that handles all the escaped local variables and used |
| 468 | /// instead of their original storage. |
| 469 | const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) { |
| 470 | if (!GlobalizedRD) |
| 471 | buildRecordForGlobalizedVars(IsInTTDRegion); |
| 472 | return GlobalizedRD; |
| 473 | } |
| 474 | |
| 475 | /// Returns the field in the globalized record for the escaped variable. |
| 476 | const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const { |
| 477 | assert(GlobalizedRD && |
| 478 | "Record for globalized variables must be generated already." ); |
| 479 | return MappedDeclsFields.lookup(Val: VD); |
| 480 | } |
| 481 | |
| 482 | /// Returns the list of the escaped local variables/parameters. |
| 483 | ArrayRef<const ValueDecl *> getEscapedDecls() const { |
| 484 | return EscapedDecls.getArrayRef(); |
| 485 | } |
| 486 | |
| 487 | /// Checks if the escaped local variable is actually a parameter passed by |
| 488 | /// value. |
| 489 | const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const { |
| 490 | return EscapedParameters; |
| 491 | } |
| 492 | |
| 493 | /// Returns the list of the escaped variables with the variably modified |
| 494 | /// types. |
| 495 | ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const { |
| 496 | return EscapedVariableLengthDecls.getArrayRef(); |
| 497 | } |
| 498 | |
| 499 | /// Returns the list of the delayed variables with the variably modified |
| 500 | /// types. |
| 501 | ArrayRef<const ValueDecl *> getDelayedVariableLengthDecls() const { |
| 502 | return DelayedVariableLengthDecls.getArrayRef(); |
| 503 | } |
| 504 | }; |
| 505 | } // anonymous namespace |
| 506 | |
| 507 | CGOpenMPRuntimeGPU::ExecutionMode |
| 508 | CGOpenMPRuntimeGPU::getExecutionMode() const { |
| 509 | return CurrentExecutionMode; |
| 510 | } |
| 511 | |
| 512 | CGOpenMPRuntimeGPU::DataSharingMode |
| 513 | CGOpenMPRuntimeGPU::getDataSharingMode() const { |
| 514 | return CurrentDataSharingMode; |
| 515 | } |
| 516 | |
| 517 | /// Check for inner (nested) SPMD construct, if any |
| 518 | static bool hasNestedSPMDDirective(ASTContext &Ctx, |
| 519 | const OMPExecutableDirective &D) { |
| 520 | const auto *CS = D.getInnermostCapturedStmt(); |
| 521 | const auto *Body = |
| 522 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); |
| 523 | const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 524 | |
| 525 | if (const auto *NestedDir = |
| 526 | dyn_cast_or_null<OMPExecutableDirective>(Val: ChildStmt)) { |
| 527 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); |
| 528 | switch (D.getDirectiveKind()) { |
| 529 | case OMPD_target: |
| 530 | if (isOpenMPParallelDirective(DKind)) |
| 531 | return true; |
| 532 | if (DKind == OMPD_teams) { |
| 533 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
| 534 | /*IgnoreCaptured=*/true); |
| 535 | if (!Body) |
| 536 | return false; |
| 537 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 538 | if (const auto *NND = |
| 539 | dyn_cast_or_null<OMPExecutableDirective>(Val: ChildStmt)) { |
| 540 | DKind = NND->getDirectiveKind(); |
| 541 | if (isOpenMPParallelDirective(DKind)) |
| 542 | return true; |
| 543 | } |
| 544 | } |
| 545 | return false; |
| 546 | case OMPD_target_teams: |
| 547 | return isOpenMPParallelDirective(DKind); |
| 548 | case OMPD_target_simd: |
| 549 | case OMPD_target_parallel: |
| 550 | case OMPD_target_parallel_for: |
| 551 | case OMPD_target_parallel_for_simd: |
| 552 | case OMPD_target_teams_distribute: |
| 553 | case OMPD_target_teams_distribute_simd: |
| 554 | case OMPD_target_teams_distribute_parallel_for: |
| 555 | case OMPD_target_teams_distribute_parallel_for_simd: |
| 556 | case OMPD_parallel: |
| 557 | case OMPD_for: |
| 558 | case OMPD_parallel_for: |
| 559 | case OMPD_parallel_master: |
| 560 | case OMPD_parallel_sections: |
| 561 | case OMPD_for_simd: |
| 562 | case OMPD_parallel_for_simd: |
| 563 | case OMPD_cancel: |
| 564 | case OMPD_cancellation_point: |
| 565 | case OMPD_ordered_standalone: |
| 566 | case OMPD_ordered_blockassoc: |
| 567 | case OMPD_threadprivate: |
| 568 | case OMPD_allocate: |
| 569 | case OMPD_task: |
| 570 | case OMPD_simd: |
| 571 | case OMPD_sections: |
| 572 | case OMPD_section: |
| 573 | case OMPD_single: |
| 574 | case OMPD_master: |
| 575 | case OMPD_critical: |
| 576 | case OMPD_taskyield: |
| 577 | case OMPD_barrier: |
| 578 | case OMPD_taskwait: |
| 579 | case OMPD_taskgroup: |
| 580 | case OMPD_atomic: |
| 581 | case OMPD_flush: |
| 582 | case OMPD_depobj: |
| 583 | case OMPD_scan: |
| 584 | case OMPD_teams: |
| 585 | case OMPD_target_data: |
| 586 | case OMPD_target_exit_data: |
| 587 | case OMPD_target_enter_data: |
| 588 | case OMPD_distribute: |
| 589 | case OMPD_distribute_simd: |
| 590 | case OMPD_distribute_parallel_for: |
| 591 | case OMPD_distribute_parallel_for_simd: |
| 592 | case OMPD_teams_distribute: |
| 593 | case OMPD_teams_distribute_simd: |
| 594 | case OMPD_teams_distribute_parallel_for: |
| 595 | case OMPD_teams_distribute_parallel_for_simd: |
| 596 | case OMPD_target_update: |
| 597 | case OMPD_declare_simd: |
| 598 | case OMPD_declare_variant: |
| 599 | case OMPD_begin_declare_variant: |
| 600 | case OMPD_end_declare_variant: |
| 601 | case OMPD_declare_target: |
| 602 | case OMPD_end_declare_target: |
| 603 | case OMPD_declare_reduction: |
| 604 | case OMPD_declare_mapper: |
| 605 | case OMPD_taskloop: |
| 606 | case OMPD_taskloop_simd: |
| 607 | case OMPD_master_taskloop: |
| 608 | case OMPD_master_taskloop_simd: |
| 609 | case OMPD_parallel_master_taskloop: |
| 610 | case OMPD_parallel_master_taskloop_simd: |
| 611 | case OMPD_requires: |
| 612 | case OMPD_unknown: |
| 613 | default: |
| 614 | llvm_unreachable("Unexpected directive." ); |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | return false; |
| 619 | } |
| 620 | |
| 621 | static bool supportsSPMDExecutionMode(ASTContext &Ctx, |
| 622 | const OMPExecutableDirective &D) { |
| 623 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); |
| 624 | switch (DirectiveKind) { |
| 625 | case OMPD_target: |
| 626 | case OMPD_target_teams: |
| 627 | return hasNestedSPMDDirective(Ctx, D); |
| 628 | case OMPD_target_parallel_loop: |
| 629 | case OMPD_target_parallel: |
| 630 | case OMPD_target_parallel_for: |
| 631 | case OMPD_target_parallel_for_simd: |
| 632 | case OMPD_target_teams_distribute_parallel_for: |
| 633 | case OMPD_target_teams_distribute_parallel_for_simd: |
| 634 | case OMPD_target_simd: |
| 635 | case OMPD_target_teams_distribute_simd: |
| 636 | return true; |
| 637 | case OMPD_target_teams_distribute: |
| 638 | return false; |
| 639 | case OMPD_target_teams_loop: |
| 640 | // Whether this is true or not depends on how the directive will |
| 641 | // eventually be emitted. |
| 642 | if (auto *TTLD = dyn_cast<OMPTargetTeamsGenericLoopDirective>(Val: &D)) |
| 643 | return TTLD->canBeParallelFor(); |
| 644 | return false; |
| 645 | case OMPD_parallel: |
| 646 | case OMPD_for: |
| 647 | case OMPD_parallel_for: |
| 648 | case OMPD_parallel_master: |
| 649 | case OMPD_parallel_sections: |
| 650 | case OMPD_for_simd: |
| 651 | case OMPD_parallel_for_simd: |
| 652 | case OMPD_cancel: |
| 653 | case OMPD_cancellation_point: |
| 654 | case OMPD_ordered_standalone: |
| 655 | case OMPD_ordered_blockassoc: |
| 656 | case OMPD_threadprivate: |
| 657 | case OMPD_allocate: |
| 658 | case OMPD_task: |
| 659 | case OMPD_simd: |
| 660 | case OMPD_sections: |
| 661 | case OMPD_section: |
| 662 | case OMPD_single: |
| 663 | case OMPD_master: |
| 664 | case OMPD_critical: |
| 665 | case OMPD_taskyield: |
| 666 | case OMPD_barrier: |
| 667 | case OMPD_taskwait: |
| 668 | case OMPD_taskgroup: |
| 669 | case OMPD_atomic: |
| 670 | case OMPD_flush: |
| 671 | case OMPD_depobj: |
| 672 | case OMPD_scan: |
| 673 | case OMPD_teams: |
| 674 | case OMPD_target_data: |
| 675 | case OMPD_target_exit_data: |
| 676 | case OMPD_target_enter_data: |
| 677 | case OMPD_distribute: |
| 678 | case OMPD_distribute_simd: |
| 679 | case OMPD_distribute_parallel_for: |
| 680 | case OMPD_distribute_parallel_for_simd: |
| 681 | case OMPD_teams_distribute: |
| 682 | case OMPD_teams_distribute_simd: |
| 683 | case OMPD_teams_distribute_parallel_for: |
| 684 | case OMPD_teams_distribute_parallel_for_simd: |
| 685 | case OMPD_target_update: |
| 686 | case OMPD_declare_simd: |
| 687 | case OMPD_declare_variant: |
| 688 | case OMPD_begin_declare_variant: |
| 689 | case OMPD_end_declare_variant: |
| 690 | case OMPD_declare_target: |
| 691 | case OMPD_end_declare_target: |
| 692 | case OMPD_declare_reduction: |
| 693 | case OMPD_declare_mapper: |
| 694 | case OMPD_taskloop: |
| 695 | case OMPD_taskloop_simd: |
| 696 | case OMPD_master_taskloop: |
| 697 | case OMPD_master_taskloop_simd: |
| 698 | case OMPD_parallel_master_taskloop: |
| 699 | case OMPD_parallel_master_taskloop_simd: |
| 700 | case OMPD_requires: |
| 701 | case OMPD_unknown: |
| 702 | default: |
| 703 | break; |
| 704 | } |
| 705 | llvm_unreachable( |
| 706 | "Unknown programming model for OpenMP directive on NVPTX target." ); |
| 707 | } |
| 708 | |
| 709 | void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D, |
| 710 | StringRef ParentName, |
| 711 | llvm::Function *&OutlinedFn, |
| 712 | llvm::Constant *&OutlinedFnID, |
| 713 | bool IsOffloadEntry, |
| 714 | const RegionCodeGenTy &CodeGen) { |
| 715 | ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_NonSPMD); |
| 716 | EntryFunctionState EST; |
| 717 | WrapperFunctionsMap.clear(); |
| 718 | |
| 719 | [[maybe_unused]] bool IsBareKernel = D.getSingleClause<OMPXBareClause>(); |
| 720 | assert(!IsBareKernel && "bare kernel should not be at generic mode" ); |
| 721 | |
| 722 | // Emit target region as a standalone region. |
| 723 | class NVPTXPrePostActionTy : public PrePostActionTy { |
| 724 | CGOpenMPRuntimeGPU::EntryFunctionState &EST; |
| 725 | const OMPExecutableDirective &D; |
| 726 | |
| 727 | public: |
| 728 | NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST, |
| 729 | const OMPExecutableDirective &D) |
| 730 | : EST(EST), D(D) {} |
| 731 | void Enter(CodeGenFunction &CGF) override { |
| 732 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 733 | RT.emitKernelInit(D, CGF, EST, /* IsSPMD */ false); |
| 734 | // Skip target region initialization. |
| 735 | RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); |
| 736 | } |
| 737 | void Exit(CodeGenFunction &CGF) override { |
| 738 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 739 | RT.clearLocThreadIdInsertPt(CGF); |
| 740 | RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false); |
| 741 | } |
| 742 | } Action(EST, D); |
| 743 | CodeGen.setAction(Action); |
| 744 | IsInTTDRegion = true; |
| 745 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, |
| 746 | IsOffloadEntry, CodeGen); |
| 747 | IsInTTDRegion = false; |
| 748 | } |
| 749 | |
| 750 | void CGOpenMPRuntimeGPU::emitKernelInit(const OMPExecutableDirective &D, |
| 751 | CodeGenFunction &CGF, |
| 752 | EntryFunctionState &EST, bool IsSPMD) { |
| 753 | llvm::OpenMPIRBuilder::TargetKernelDefaultAttrs Attrs; |
| 754 | Attrs.ExecFlags = |
| 755 | IsSPMD ? llvm::omp::OMPTgtExecModeFlags::OMP_TGT_EXEC_MODE_SPMD |
| 756 | : llvm::omp::OMPTgtExecModeFlags::OMP_TGT_EXEC_MODE_GENERIC; |
| 757 | computeMinAndMaxThreadsAndTeams(D, CGF, Attrs); |
| 758 | |
| 759 | CGBuilderTy &Bld = CGF.Builder; |
| 760 | Bld.restoreIP(IP: OMPBuilder.createTargetInit(Loc: Bld, Attrs)); |
| 761 | if (!IsSPMD) |
| 762 | emitGenericVarsProlog(CGF, Loc: EST.Loc); |
| 763 | } |
| 764 | |
| 765 | void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF, |
| 766 | EntryFunctionState &EST, |
| 767 | bool IsSPMD) { |
| 768 | if (!IsSPMD) |
| 769 | emitGenericVarsEpilog(CGF); |
| 770 | |
| 771 | // This is temporary until we remove the fixed sized buffer. |
| 772 | ASTContext &C = CGM.getContext(); |
| 773 | RecordDecl *StaticRD = C.buildImplicitRecord( |
| 774 | Name: "_openmp_teams_reduction_type_$_" , TK: RecordDecl::TagKind::Union); |
| 775 | StaticRD->startDefinition(); |
| 776 | for (const RecordDecl *TeamReductionRec : TeamsReductions) { |
| 777 | CanQualType RecTy = C.getCanonicalTagType(TD: TeamReductionRec); |
| 778 | auto *Field = FieldDecl::Create( |
| 779 | C, DC: StaticRD, StartLoc: SourceLocation(), IdLoc: SourceLocation(), Id: nullptr, T: RecTy, |
| 780 | TInfo: C.getTrivialTypeSourceInfo(T: RecTy, Loc: SourceLocation()), |
| 781 | /*BW=*/nullptr, /*Mutable=*/false, |
| 782 | /*InitStyle=*/ICIS_NoInit); |
| 783 | Field->setAccess(AS_public); |
| 784 | StaticRD->addDecl(D: Field); |
| 785 | } |
| 786 | StaticRD->completeDefinition(); |
| 787 | CanQualType StaticTy = C.getCanonicalTagType(TD: StaticRD); |
| 788 | llvm::Type *LLVMReductionsBufferTy = |
| 789 | CGM.getTypes().ConvertTypeForMem(T: StaticTy); |
| 790 | const auto &DL = CGM.getModule().getDataLayout(); |
| 791 | uint64_t ReductionDataSize = |
| 792 | TeamsReductions.empty() |
| 793 | ? 0 |
| 794 | : DL.getTypeAllocSize(Ty: LLVMReductionsBufferTy).getFixedValue(); |
| 795 | CGBuilderTy &Bld = CGF.Builder; |
| 796 | OMPBuilder.createTargetDeinit(Loc: Bld, TeamsReductionDataSize: ReductionDataSize); |
| 797 | TeamsReductions.clear(); |
| 798 | } |
| 799 | |
| 800 | void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D, |
| 801 | StringRef ParentName, |
| 802 | llvm::Function *&OutlinedFn, |
| 803 | llvm::Constant *&OutlinedFnID, |
| 804 | bool IsOffloadEntry, |
| 805 | const RegionCodeGenTy &CodeGen) { |
| 806 | ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_SPMD); |
| 807 | EntryFunctionState EST; |
| 808 | |
| 809 | bool IsBareKernel = D.getSingleClause<OMPXBareClause>(); |
| 810 | |
| 811 | // Emit target region as a standalone region. |
| 812 | class NVPTXPrePostActionTy : public PrePostActionTy { |
| 813 | CGOpenMPRuntimeGPU &RT; |
| 814 | CGOpenMPRuntimeGPU::EntryFunctionState &EST; |
| 815 | bool IsBareKernel; |
| 816 | DataSharingMode Mode; |
| 817 | const OMPExecutableDirective &D; |
| 818 | |
| 819 | public: |
| 820 | NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT, |
| 821 | CGOpenMPRuntimeGPU::EntryFunctionState &EST, |
| 822 | bool IsBareKernel, const OMPExecutableDirective &D) |
| 823 | : RT(RT), EST(EST), IsBareKernel(IsBareKernel), |
| 824 | Mode(RT.CurrentDataSharingMode), D(D) {} |
| 825 | void Enter(CodeGenFunction &CGF) override { |
| 826 | if (IsBareKernel) { |
| 827 | RT.CurrentDataSharingMode = DataSharingMode::DS_CUDA; |
| 828 | return; |
| 829 | } |
| 830 | RT.emitKernelInit(D, CGF, EST, /* IsSPMD */ true); |
| 831 | // Skip target region initialization. |
| 832 | RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); |
| 833 | } |
| 834 | void Exit(CodeGenFunction &CGF) override { |
| 835 | if (IsBareKernel) { |
| 836 | RT.CurrentDataSharingMode = Mode; |
| 837 | return; |
| 838 | } |
| 839 | RT.clearLocThreadIdInsertPt(CGF); |
| 840 | RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true); |
| 841 | } |
| 842 | } Action(*this, EST, IsBareKernel, D); |
| 843 | CodeGen.setAction(Action); |
| 844 | IsInTTDRegion = true; |
| 845 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, |
| 846 | IsOffloadEntry, CodeGen); |
| 847 | IsInTTDRegion = false; |
| 848 | } |
| 849 | |
| 850 | void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction( |
| 851 | const OMPExecutableDirective &D, StringRef ParentName, |
| 852 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, |
| 853 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { |
| 854 | if (!IsOffloadEntry) // Nothing to do. |
| 855 | return; |
| 856 | |
| 857 | assert(!ParentName.empty() && "Invalid target region parent name!" ); |
| 858 | |
| 859 | bool Mode = supportsSPMDExecutionMode(Ctx&: CGM.getContext(), D); |
| 860 | bool IsBareKernel = D.getSingleClause<OMPXBareClause>(); |
| 861 | if (Mode || IsBareKernel) |
| 862 | emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, |
| 863 | CodeGen); |
| 864 | else |
| 865 | emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, |
| 866 | CodeGen); |
| 867 | } |
| 868 | |
| 869 | CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM) |
| 870 | : CGOpenMPRuntime(CGM) { |
| 871 | llvm::OpenMPIRBuilderConfig Config( |
| 872 | CGM.getLangOpts().OpenMPIsTargetDevice, isGPU(), |
| 873 | CGM.getLangOpts().OpenMPOffloadMandatory, |
| 874 | /*HasRequiresReverseOffload*/ false, /*HasRequiresUnifiedAddress*/ false, |
| 875 | hasRequiresUnifiedSharedMemory(), /*HasRequiresDynamicAllocators*/ false); |
| 876 | Config.setDefaultTargetAS( |
| 877 | CGM.getContext().getTargetInfo().getTargetAddressSpace(AS: LangAS::Default)); |
| 878 | Config.setRuntimeCC(CGM.getRuntimeCC()); |
| 879 | |
| 880 | OMPBuilder.setConfig(Config); |
| 881 | |
| 882 | if (!CGM.getLangOpts().OpenMPIsTargetDevice) |
| 883 | llvm_unreachable("OpenMP can only handle device code." ); |
| 884 | |
| 885 | if (CGM.getLangOpts().OpenMPCUDAMode) |
| 886 | CurrentDataSharingMode = CGOpenMPRuntimeGPU::DS_CUDA; |
| 887 | |
| 888 | llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder(); |
| 889 | if (CGM.getLangOpts().NoGPULib || CGM.getLangOpts().OMPHostIRFile.empty()) |
| 890 | return; |
| 891 | |
| 892 | OMPBuilder.createGlobalFlag(Value: CGM.getLangOpts().OpenMPTargetDebug, |
| 893 | Name: "__omp_rtl_debug_kind" ); |
| 894 | OMPBuilder.createGlobalFlag(Value: CGM.getLangOpts().OpenMPTeamSubscription, |
| 895 | Name: "__omp_rtl_assume_teams_oversubscription" ); |
| 896 | OMPBuilder.createGlobalFlag(Value: CGM.getLangOpts().OpenMPThreadSubscription, |
| 897 | Name: "__omp_rtl_assume_threads_oversubscription" ); |
| 898 | OMPBuilder.createGlobalFlag(Value: CGM.getLangOpts().OpenMPNoThreadState, |
| 899 | Name: "__omp_rtl_assume_no_thread_state" ); |
| 900 | OMPBuilder.createGlobalFlag(Value: CGM.getLangOpts().OpenMPNoNestedParallelism, |
| 901 | Name: "__omp_rtl_assume_no_nested_parallelism" ); |
| 902 | } |
| 903 | |
| 904 | void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF, |
| 905 | ProcBindKind ProcBind, |
| 906 | SourceLocation Loc) { |
| 907 | // Nothing to do. |
| 908 | } |
| 909 | |
| 910 | llvm::Value *CGOpenMPRuntimeGPU::emitMessageClause(CodeGenFunction &CGF, |
| 911 | const Expr *Message, |
| 912 | SourceLocation Loc) { |
| 913 | CGM.getDiags().Report(Loc, DiagID: diag::warn_omp_gpu_unsupported_clause) |
| 914 | << getOpenMPClauseName(C: OMPC_message); |
| 915 | return nullptr; |
| 916 | } |
| 917 | |
| 918 | llvm::Value * |
| 919 | CGOpenMPRuntimeGPU::emitSeverityClause(OpenMPSeverityClauseKind Severity, |
| 920 | SourceLocation Loc) { |
| 921 | CGM.getDiags().Report(Loc, DiagID: diag::warn_omp_gpu_unsupported_clause) |
| 922 | << getOpenMPClauseName(C: OMPC_severity); |
| 923 | return nullptr; |
| 924 | } |
| 925 | |
| 926 | void CGOpenMPRuntimeGPU::emitNumThreadsClause( |
| 927 | CodeGenFunction &CGF, llvm::Value *NumThreads, SourceLocation Loc, |
| 928 | OpenMPNumThreadsClauseModifier Modifier, OpenMPSeverityClauseKind Severity, |
| 929 | SourceLocation SeverityLoc, const Expr *Message, |
| 930 | SourceLocation MessageLoc) { |
| 931 | if (Modifier == OMPC_NUMTHREADS_strict) { |
| 932 | CGM.getDiags().Report(Loc, |
| 933 | DiagID: diag::warn_omp_gpu_unsupported_modifier_for_clause) |
| 934 | << "strict" << getOpenMPClauseName(C: OMPC_num_threads); |
| 935 | return; |
| 936 | } |
| 937 | |
| 938 | // Nothing to do. |
| 939 | } |
| 940 | |
| 941 | void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF, |
| 942 | const Expr *NumTeams, |
| 943 | const Expr *ThreadLimit, |
| 944 | SourceLocation Loc) {} |
| 945 | |
| 946 | llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction( |
| 947 | CodeGenFunction &CGF, const OMPExecutableDirective &D, |
| 948 | const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind, |
| 949 | const RegionCodeGenTy &CodeGen) { |
| 950 | // Emit target region as a standalone region. |
| 951 | bool PrevIsInTTDRegion = IsInTTDRegion; |
| 952 | IsInTTDRegion = false; |
| 953 | auto *OutlinedFun = |
| 954 | cast<llvm::Function>(Val: CGOpenMPRuntime::emitParallelOutlinedFunction( |
| 955 | CGF, D, ThreadIDVar, InnermostKind, CodeGen)); |
| 956 | IsInTTDRegion = PrevIsInTTDRegion; |
| 957 | if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) { |
| 958 | llvm::Function *WrapperFun = |
| 959 | createParallelDataSharingWrapper(OutlinedParallelFn: OutlinedFun, D); |
| 960 | WrapperFunctionsMap[OutlinedFun] = WrapperFun; |
| 961 | } |
| 962 | |
| 963 | return OutlinedFun; |
| 964 | } |
| 965 | |
| 966 | /// Get list of lastprivate variables from the teams distribute ... or |
| 967 | /// teams {distribute ...} directives. |
| 968 | static void |
| 969 | getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D, |
| 970 | llvm::SmallVectorImpl<const ValueDecl *> &Vars) { |
| 971 | assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && |
| 972 | "expected teams directive." ); |
| 973 | const OMPExecutableDirective *Dir = &D; |
| 974 | if (!isOpenMPDistributeDirective(DKind: D.getDirectiveKind())) { |
| 975 | if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild( |
| 976 | Ctx, |
| 977 | Body: D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers( |
| 978 | /*IgnoreCaptured=*/true))) { |
| 979 | Dir = dyn_cast_or_null<OMPExecutableDirective>(Val: S); |
| 980 | if (Dir && !isOpenMPDistributeDirective(DKind: Dir->getDirectiveKind())) |
| 981 | Dir = nullptr; |
| 982 | } |
| 983 | } |
| 984 | if (!Dir) |
| 985 | return; |
| 986 | for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) { |
| 987 | for (const Expr *E : C->getVarRefs()) |
| 988 | Vars.push_back(Elt: getPrivateItem(RefExpr: E)); |
| 989 | } |
| 990 | } |
| 991 | |
| 992 | /// Get list of reduction variables from the teams ... directives. |
| 993 | static void |
| 994 | getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D, |
| 995 | llvm::SmallVectorImpl<const ValueDecl *> &Vars) { |
| 996 | assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && |
| 997 | "expected teams directive." ); |
| 998 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
| 999 | for (const Expr *E : C->privates()) |
| 1000 | Vars.push_back(Elt: getPrivateItem(RefExpr: E)); |
| 1001 | } |
| 1002 | } |
| 1003 | |
| 1004 | llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction( |
| 1005 | CodeGenFunction &CGF, const OMPExecutableDirective &D, |
| 1006 | const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind, |
| 1007 | const RegionCodeGenTy &CodeGen) { |
| 1008 | SourceLocation Loc = D.getBeginLoc(); |
| 1009 | |
| 1010 | const RecordDecl *GlobalizedRD = nullptr; |
| 1011 | llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions; |
| 1012 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; |
| 1013 | unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size; |
| 1014 | // Globalize team reductions variable unconditionally in all modes. |
| 1015 | if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) |
| 1016 | getTeamsReductionVars(Ctx&: CGM.getContext(), D, Vars&: LastPrivatesReductions); |
| 1017 | if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { |
| 1018 | getDistributeLastprivateVars(Ctx&: CGM.getContext(), D, Vars&: LastPrivatesReductions); |
| 1019 | if (!LastPrivatesReductions.empty()) { |
| 1020 | GlobalizedRD = ::buildRecordForGlobalizedVars( |
| 1021 | C&: CGM.getContext(), EscapedDecls: {}, EscapedDeclsForTeams: LastPrivatesReductions, MappedDeclsFields, |
| 1022 | BufSize: WarpSize); |
| 1023 | } |
| 1024 | } else if (!LastPrivatesReductions.empty()) { |
| 1025 | assert(!TeamAndReductions.first && |
| 1026 | "Previous team declaration is not expected." ); |
| 1027 | TeamAndReductions.first = D.getCapturedStmt(RegionKind: OMPD_teams)->getCapturedDecl(); |
| 1028 | std::swap(LHS&: TeamAndReductions.second, RHS&: LastPrivatesReductions); |
| 1029 | } |
| 1030 | |
| 1031 | // Emit target region as a standalone region. |
| 1032 | class NVPTXPrePostActionTy : public PrePostActionTy { |
| 1033 | SourceLocation &Loc; |
| 1034 | const RecordDecl *GlobalizedRD; |
| 1035 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 1036 | &MappedDeclsFields; |
| 1037 | |
| 1038 | public: |
| 1039 | NVPTXPrePostActionTy( |
| 1040 | SourceLocation &Loc, const RecordDecl *GlobalizedRD, |
| 1041 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 1042 | &MappedDeclsFields) |
| 1043 | : Loc(Loc), GlobalizedRD(GlobalizedRD), |
| 1044 | MappedDeclsFields(MappedDeclsFields) {} |
| 1045 | void Enter(CodeGenFunction &CGF) override { |
| 1046 | auto &Rt = |
| 1047 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 1048 | if (GlobalizedRD) { |
| 1049 | auto I = Rt.FunctionGlobalizedDecls.try_emplace(Key: CGF.CurFn).first; |
| 1050 | I->getSecond().MappedParams = |
| 1051 | std::make_unique<CodeGenFunction::OMPMapVars>(); |
| 1052 | DeclToAddrMapTy &Data = I->getSecond().LocalVarData; |
| 1053 | for (const auto &Pair : MappedDeclsFields) { |
| 1054 | assert(Pair.getFirst()->isCanonicalDecl() && |
| 1055 | "Expected canonical declaration" ); |
| 1056 | Data.try_emplace(Key: Pair.getFirst()); |
| 1057 | } |
| 1058 | } |
| 1059 | Rt.emitGenericVarsProlog(CGF, Loc); |
| 1060 | } |
| 1061 | void Exit(CodeGenFunction &CGF) override { |
| 1062 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) |
| 1063 | .emitGenericVarsEpilog(CGF); |
| 1064 | } |
| 1065 | } Action(Loc, GlobalizedRD, MappedDeclsFields); |
| 1066 | CodeGen.setAction(Action); |
| 1067 | llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction( |
| 1068 | CGF, D, ThreadIDVar, InnermostKind, CodeGen); |
| 1069 | |
| 1070 | return OutlinedFun; |
| 1071 | } |
| 1072 | |
| 1073 | void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF, |
| 1074 | SourceLocation Loc) { |
| 1075 | if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic) |
| 1076 | return; |
| 1077 | |
| 1078 | CGBuilderTy &Bld = CGF.Builder; |
| 1079 | |
| 1080 | const auto I = FunctionGlobalizedDecls.find(Val: CGF.CurFn); |
| 1081 | if (I == FunctionGlobalizedDecls.end()) |
| 1082 | return; |
| 1083 | |
| 1084 | for (auto &Rec : I->getSecond().LocalVarData) { |
| 1085 | const auto *VD = cast<VarDecl>(Val: Rec.first); |
| 1086 | bool EscapedParam = I->getSecond().EscapedParameters.count(Ptr: Rec.first); |
| 1087 | QualType VarTy = VD->getType(); |
| 1088 | |
| 1089 | // Get the local allocation of a firstprivate variable before sharing |
| 1090 | llvm::Value *ParValue; |
| 1091 | if (EscapedParam) { |
| 1092 | LValue ParLVal = |
| 1093 | CGF.MakeAddrLValue(Addr: CGF.GetAddrOfLocalVar(VD), T: VD->getType()); |
| 1094 | ParValue = CGF.EmitLoadOfScalar(lvalue: ParLVal, Loc); |
| 1095 | } |
| 1096 | |
| 1097 | // Allocate space for the variable to be globalized |
| 1098 | llvm::Value *AllocArgs[] = {CGF.getTypeSize(Ty: VD->getType())}; |
| 1099 | llvm::CallBase *VoidPtr = |
| 1100 | CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1101 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_alloc_shared), |
| 1102 | args: AllocArgs, name: VD->getName()); |
| 1103 | // FIXME: We should use the variables actual alignment as an argument. |
| 1104 | VoidPtr->addRetAttr(Attr: llvm::Attribute::get( |
| 1105 | Context&: CGM.getLLVMContext(), Kind: llvm::Attribute::Alignment, |
| 1106 | Val: CGM.getContext().getTargetInfo().getNewAlign() / 8)); |
| 1107 | |
| 1108 | // Cast the void pointer and get the address of the globalized variable. |
| 1109 | llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 1110 | V: VoidPtr, DestTy: Bld.getPtrTy(AddrSpace: 0), Name: VD->getName() + "_on_stack" ); |
| 1111 | LValue VarAddr = |
| 1112 | CGF.MakeNaturalAlignPointeeRawAddrLValue(V: CastedVoidPtr, T: VarTy); |
| 1113 | Rec.second.PrivateAddr = VarAddr.getAddress(); |
| 1114 | Rec.second.GlobalizedVal = VoidPtr; |
| 1115 | |
| 1116 | // Assign the local allocation to the newly globalized location. |
| 1117 | if (EscapedParam) { |
| 1118 | CGF.EmitStoreOfScalar(value: ParValue, lvalue: VarAddr); |
| 1119 | I->getSecond().MappedParams->setVarAddr(CGF, LocalVD: VD, TempAddr: VarAddr.getAddress()); |
| 1120 | } |
| 1121 | if (auto *DI = CGF.getDebugInfo()) |
| 1122 | VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(Loc: VD->getLocation())); |
| 1123 | } |
| 1124 | |
| 1125 | for (const auto *ValueD : I->getSecond().EscapedVariableLengthDecls) { |
| 1126 | const auto *VD = cast<VarDecl>(Val: ValueD); |
| 1127 | std::pair<llvm::Value *, llvm::Value *> AddrSizePair = |
| 1128 | getKmpcAllocShared(CGF, VD); |
| 1129 | I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(Args&: AddrSizePair); |
| 1130 | LValue Base = CGF.MakeAddrLValue(V: AddrSizePair.first, T: VD->getType(), |
| 1131 | Alignment: CGM.getContext().getDeclAlign(D: VD), |
| 1132 | Source: AlignmentSource::Decl); |
| 1133 | I->getSecond().MappedParams->setVarAddr(CGF, LocalVD: VD, TempAddr: Base.getAddress()); |
| 1134 | } |
| 1135 | I->getSecond().MappedParams->apply(CGF); |
| 1136 | } |
| 1137 | |
| 1138 | bool CGOpenMPRuntimeGPU::isDelayedVariableLengthDecl(CodeGenFunction &CGF, |
| 1139 | const VarDecl *VD) const { |
| 1140 | const auto I = FunctionGlobalizedDecls.find(Val: CGF.CurFn); |
| 1141 | if (I == FunctionGlobalizedDecls.end()) |
| 1142 | return false; |
| 1143 | |
| 1144 | // Check variable declaration is delayed: |
| 1145 | return llvm::is_contained(Range: I->getSecond().DelayedVariableLengthDecls, Element: VD); |
| 1146 | } |
| 1147 | |
| 1148 | std::pair<llvm::Value *, llvm::Value *> |
| 1149 | CGOpenMPRuntimeGPU::getKmpcAllocShared(CodeGenFunction &CGF, |
| 1150 | const VarDecl *VD) { |
| 1151 | CGBuilderTy &Bld = CGF.Builder; |
| 1152 | |
| 1153 | // Compute size and alignment. |
| 1154 | llvm::Value *Size = CGF.getTypeSize(Ty: VD->getType()); |
| 1155 | CharUnits Align = CGM.getContext().getDeclAlign(D: VD); |
| 1156 | Size = Bld.CreateNUWAdd( |
| 1157 | LHS: Size, RHS: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: Align.getQuantity() - 1)); |
| 1158 | llvm::Value *AlignVal = |
| 1159 | llvm::ConstantInt::get(Ty: CGF.SizeTy, V: Align.getQuantity()); |
| 1160 | Size = Bld.CreateUDiv(LHS: Size, RHS: AlignVal); |
| 1161 | Size = Bld.CreateNUWMul(LHS: Size, RHS: AlignVal); |
| 1162 | |
| 1163 | // Allocate space for this VLA object to be globalized. |
| 1164 | llvm::Value *AllocArgs[] = {Size}; |
| 1165 | llvm::CallBase *VoidPtr = |
| 1166 | CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1167 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_alloc_shared), |
| 1168 | args: AllocArgs, name: VD->getName()); |
| 1169 | VoidPtr->addRetAttr(Attr: llvm::Attribute::get( |
| 1170 | Context&: CGM.getLLVMContext(), Kind: llvm::Attribute::Alignment, Val: Align.getQuantity())); |
| 1171 | |
| 1172 | return std::make_pair(x&: VoidPtr, y&: Size); |
| 1173 | } |
| 1174 | |
| 1175 | void CGOpenMPRuntimeGPU::getKmpcFreeShared( |
| 1176 | CodeGenFunction &CGF, |
| 1177 | const std::pair<llvm::Value *, llvm::Value *> &AddrSizePair) { |
| 1178 | // Deallocate the memory for each globalized VLA object |
| 1179 | CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1180 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_free_shared), |
| 1181 | args: {AddrSizePair.first, AddrSizePair.second}); |
| 1182 | } |
| 1183 | |
| 1184 | void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF) { |
| 1185 | if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic) |
| 1186 | return; |
| 1187 | |
| 1188 | const auto I = FunctionGlobalizedDecls.find(Val: CGF.CurFn); |
| 1189 | if (I != FunctionGlobalizedDecls.end()) { |
| 1190 | // Deallocate the memory for each globalized VLA object that was |
| 1191 | // globalized in the prolog (i.e. emitGenericVarsProlog). |
| 1192 | for (const auto &AddrSizePair : |
| 1193 | llvm::reverse(C&: I->getSecond().EscapedVariableLengthDeclsAddrs)) { |
| 1194 | CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1195 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_free_shared), |
| 1196 | args: {AddrSizePair.first, AddrSizePair.second}); |
| 1197 | } |
| 1198 | // Deallocate the memory for each globalized value |
| 1199 | for (auto &Rec : llvm::reverse(C&: I->getSecond().LocalVarData)) { |
| 1200 | const auto *VD = cast<VarDecl>(Val: Rec.first); |
| 1201 | I->getSecond().MappedParams->restore(CGF); |
| 1202 | |
| 1203 | llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal, |
| 1204 | CGF.getTypeSize(Ty: VD->getType())}; |
| 1205 | CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1206 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_free_shared), |
| 1207 | args: FreeArgs); |
| 1208 | } |
| 1209 | } |
| 1210 | } |
| 1211 | |
| 1212 | void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF, |
| 1213 | const OMPExecutableDirective &D, |
| 1214 | SourceLocation Loc, |
| 1215 | llvm::Function *OutlinedFn, |
| 1216 | ArrayRef<llvm::Value *> CapturedVars) { |
| 1217 | if (!CGF.HaveInsertPoint()) |
| 1218 | return; |
| 1219 | |
| 1220 | bool IsBareKernel = D.getSingleClause<OMPXBareClause>(); |
| 1221 | |
| 1222 | RawAddress ZeroAddr = CGF.CreateDefaultAlignTempAlloca(Ty: CGF.Int32Ty, |
| 1223 | /*Name=*/".zero.addr" ); |
| 1224 | CGF.Builder.CreateStore(Val: CGF.Builder.getInt32(/*C*/ 0), Addr: ZeroAddr); |
| 1225 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; |
| 1226 | // We don't emit any thread id function call in bare kernel, but because the |
| 1227 | // outlined function has a pointer argument, we emit a nullptr here. |
| 1228 | if (IsBareKernel) |
| 1229 | OutlinedFnArgs.push_back(Elt: llvm::ConstantPointerNull::get(T: CGM.VoidPtrTy)); |
| 1230 | else |
| 1231 | OutlinedFnArgs.push_back(Elt: emitThreadIDAddress(CGF, Loc).emitRawPointer(CGF)); |
| 1232 | OutlinedFnArgs.push_back(Elt: ZeroAddr.getPointer()); |
| 1233 | OutlinedFnArgs.append(in_start: CapturedVars.begin(), in_end: CapturedVars.end()); |
| 1234 | emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, Args: OutlinedFnArgs); |
| 1235 | } |
| 1236 | |
| 1237 | void CGOpenMPRuntimeGPU::emitParallelCall( |
| 1238 | CodeGenFunction &CGF, SourceLocation Loc, llvm::Function *OutlinedFn, |
| 1239 | ArrayRef<llvm::Value *> CapturedVars, const Expr *IfCond, |
| 1240 | llvm::Value *NumThreads, OpenMPNumThreadsClauseModifier NumThreadsModifier, |
| 1241 | OpenMPSeverityClauseKind Severity, const Expr *Message) { |
| 1242 | if (!CGF.HaveInsertPoint()) |
| 1243 | return; |
| 1244 | |
| 1245 | auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond, |
| 1246 | NumThreads](CodeGenFunction &CGF, |
| 1247 | PrePostActionTy &Action) { |
| 1248 | CGBuilderTy &Bld = CGF.Builder; |
| 1249 | llvm::Value *NumThreadsVal = NumThreads; |
| 1250 | llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn]; |
| 1251 | llvm::PointerType *FnPtrTy = llvm::PointerType::get( |
| 1252 | C&: CGF.getLLVMContext(), AddressSpace: CGM.getDataLayout().getProgramAddressSpace()); |
| 1253 | |
| 1254 | llvm::Value *ID = llvm::ConstantPointerNull::get(T: FnPtrTy); |
| 1255 | if (WFn) |
| 1256 | ID = Bld.CreateBitOrPointerCast(V: WFn, DestTy: FnPtrTy); |
| 1257 | |
| 1258 | llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(V: OutlinedFn, DestTy: FnPtrTy); |
| 1259 | |
| 1260 | // Create a private scope that will globalize the arguments |
| 1261 | // passed from the outside of the target region. |
| 1262 | // TODO: Is that needed? |
| 1263 | CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF); |
| 1264 | |
| 1265 | Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca( |
| 1266 | Ty: llvm::ArrayType::get(ElementType: CGM.VoidPtrTy, NumElements: CapturedVars.size()), |
| 1267 | Name: "captured_vars_addrs" ); |
| 1268 | // There's something to share. |
| 1269 | if (!CapturedVars.empty()) { |
| 1270 | // Prepare for parallel region. Indicate the outlined function. |
| 1271 | ASTContext &Ctx = CGF.getContext(); |
| 1272 | unsigned Idx = 0; |
| 1273 | for (llvm::Value *V : CapturedVars) { |
| 1274 | Address Dst = Bld.CreateConstArrayGEP(Addr: CapturedVarsAddrs, Index: Idx); |
| 1275 | llvm::Value *PtrV; |
| 1276 | if (V->getType()->isIntegerTy()) |
| 1277 | PtrV = Bld.CreateIntToPtr(V, DestTy: CGF.VoidPtrTy); |
| 1278 | else |
| 1279 | PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, DestTy: CGF.VoidPtrTy); |
| 1280 | CGF.EmitStoreOfScalar(Value: PtrV, Addr: Dst, /*Volatile=*/false, |
| 1281 | Ty: Ctx.getPointerType(T: Ctx.VoidPtrTy)); |
| 1282 | ++Idx; |
| 1283 | } |
| 1284 | } |
| 1285 | |
| 1286 | llvm::Value *IfCondVal = nullptr; |
| 1287 | if (IfCond) |
| 1288 | IfCondVal = Bld.CreateIntCast(V: CGF.EvaluateExprAsBool(E: IfCond), DestTy: CGF.Int32Ty, |
| 1289 | /* isSigned */ false); |
| 1290 | else |
| 1291 | IfCondVal = llvm::ConstantInt::get(Ty: CGF.Int32Ty, V: 1); |
| 1292 | |
| 1293 | if (!NumThreadsVal) |
| 1294 | NumThreadsVal = llvm::ConstantInt::getAllOnesValue(Ty: CGF.Int32Ty); |
| 1295 | else |
| 1296 | NumThreadsVal = Bld.CreateZExtOrTrunc(V: NumThreadsVal, DestTy: CGF.Int32Ty); |
| 1297 | |
| 1298 | // No strict prescriptiveness for the number of threads. |
| 1299 | llvm::Value *StrictNumThreadsVal = llvm::ConstantInt::get(Ty: CGF.Int32Ty, V: 0); |
| 1300 | |
| 1301 | assert(IfCondVal && "Expected a value" ); |
| 1302 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); |
| 1303 | llvm::Value *Args[] = { |
| 1304 | RTLoc, |
| 1305 | getThreadID(CGF, Loc), |
| 1306 | IfCondVal, |
| 1307 | NumThreadsVal, |
| 1308 | llvm::ConstantInt::getAllOnesValue(Ty: CGF.Int32Ty), |
| 1309 | FnPtr, |
| 1310 | ID, |
| 1311 | Bld.CreateBitOrPointerCast(V: CapturedVarsAddrs.emitRawPointer(CGF), |
| 1312 | DestTy: CGF.VoidPtrPtrTy), |
| 1313 | llvm::ConstantInt::get(Ty: CGM.SizeTy, V: CapturedVars.size()), |
| 1314 | StrictNumThreadsVal}; |
| 1315 | |
| 1316 | CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1317 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_parallel_60), |
| 1318 | args: Args); |
| 1319 | }; |
| 1320 | |
| 1321 | RegionCodeGenTy RCG(ParallelGen); |
| 1322 | RCG(CGF); |
| 1323 | } |
| 1324 | |
| 1325 | void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) { |
| 1326 | // Always emit simple barriers! |
| 1327 | if (!CGF.HaveInsertPoint()) |
| 1328 | return; |
| 1329 | // Build call __kmpc_barrier_simple_spmd(nullptr, 0); |
| 1330 | // This function does not use parameters, so we can emit just default values. |
| 1331 | llvm::Value *Args[] = { |
| 1332 | llvm::ConstantPointerNull::get( |
| 1333 | T: cast<llvm::PointerType>(Val: getIdentTyPointerTy())), |
| 1334 | llvm::ConstantInt::get(Ty: CGF.Int32Ty, /*V=*/0, /*isSigned=*/IsSigned: true)}; |
| 1335 | CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1336 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_barrier_simple_spmd), |
| 1337 | args: Args); |
| 1338 | } |
| 1339 | |
| 1340 | void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF, |
| 1341 | SourceLocation Loc, |
| 1342 | OpenMPDirectiveKind Kind, bool, |
| 1343 | bool) { |
| 1344 | // Always emit simple barriers! |
| 1345 | if (!CGF.HaveInsertPoint()) |
| 1346 | return; |
| 1347 | // Build call __kmpc_cancel_barrier(loc, thread_id); |
| 1348 | unsigned Flags = getDefaultFlagsForBarriers(Kind); |
| 1349 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), |
| 1350 | getThreadID(CGF, Loc)}; |
| 1351 | |
| 1352 | CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1353 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_barrier), |
| 1354 | args: Args); |
| 1355 | } |
| 1356 | |
| 1357 | void CGOpenMPRuntimeGPU::emitCriticalRegion( |
| 1358 | CodeGenFunction &CGF, StringRef CriticalName, |
| 1359 | const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, |
| 1360 | const Expr *Hint) { |
| 1361 | llvm::BasicBlock *LoopBB = CGF.createBasicBlock(name: "omp.critical.loop" ); |
| 1362 | llvm::BasicBlock *TestBB = CGF.createBasicBlock(name: "omp.critical.test" ); |
| 1363 | llvm::BasicBlock *SyncBB = CGF.createBasicBlock(name: "omp.critical.sync" ); |
| 1364 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock(name: "omp.critical.body" ); |
| 1365 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(name: "omp.critical.exit" ); |
| 1366 | |
| 1367 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 1368 | |
| 1369 | // Get the mask of active threads in the warp. |
| 1370 | llvm::Value *Mask = CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1371 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_warp_active_thread_mask)); |
| 1372 | // Fetch team-local id of the thread. |
| 1373 | llvm::Value *ThreadID = RT.getGPUThreadID(CGF); |
| 1374 | |
| 1375 | // Get the width of the team. |
| 1376 | llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF); |
| 1377 | |
| 1378 | // Initialize the counter variable for the loop. |
| 1379 | QualType Int32Ty = |
| 1380 | CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0); |
| 1381 | Address Counter = CGF.CreateMemTempWithoutCast(T: Int32Ty, Name: "critical_counter" ); |
| 1382 | LValue CounterLVal = CGF.MakeAddrLValue(Addr: Counter, T: Int32Ty); |
| 1383 | CGF.EmitStoreOfScalar(value: llvm::Constant::getNullValue(Ty: CGM.Int32Ty), lvalue: CounterLVal, |
| 1384 | /*isInit=*/true); |
| 1385 | |
| 1386 | // Block checks if loop counter exceeds upper bound. |
| 1387 | CGF.EmitBlock(BB: LoopBB); |
| 1388 | llvm::Value *CounterVal = CGF.EmitLoadOfScalar(lvalue: CounterLVal, Loc); |
| 1389 | llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(LHS: CounterVal, RHS: TeamWidth); |
| 1390 | CGF.Builder.CreateCondBr(Cond: CmpLoopBound, True: TestBB, False: ExitBB); |
| 1391 | |
| 1392 | // Block tests which single thread should execute region, and which threads |
| 1393 | // should go straight to synchronisation point. |
| 1394 | CGF.EmitBlock(BB: TestBB); |
| 1395 | CounterVal = CGF.EmitLoadOfScalar(lvalue: CounterLVal, Loc); |
| 1396 | llvm::Value *CmpThreadToCounter = |
| 1397 | CGF.Builder.CreateICmpEQ(LHS: ThreadID, RHS: CounterVal); |
| 1398 | CGF.Builder.CreateCondBr(Cond: CmpThreadToCounter, True: BodyBB, False: SyncBB); |
| 1399 | |
| 1400 | // Block emits the body of the critical region. |
| 1401 | CGF.EmitBlock(BB: BodyBB); |
| 1402 | |
| 1403 | // Output the critical statement. |
| 1404 | CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc, |
| 1405 | Hint); |
| 1406 | |
| 1407 | // After the body surrounded by the critical region, the single executing |
| 1408 | // thread will jump to the synchronisation point. |
| 1409 | // Block waits for all threads in current team to finish then increments the |
| 1410 | // counter variable and returns to the loop. |
| 1411 | CGF.EmitBlock(BB: SyncBB); |
| 1412 | // Reconverge active threads in the warp. |
| 1413 | (void)CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 1414 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_syncwarp), |
| 1415 | args: Mask); |
| 1416 | |
| 1417 | llvm::Value *IncCounterVal = |
| 1418 | CGF.Builder.CreateNSWAdd(LHS: CounterVal, RHS: CGF.Builder.getInt32(C: 1)); |
| 1419 | CGF.EmitStoreOfScalar(value: IncCounterVal, lvalue: CounterLVal); |
| 1420 | CGF.EmitBranch(Block: LoopBB); |
| 1421 | |
| 1422 | // Block that is reached when all threads in the team complete the region. |
| 1423 | CGF.EmitBlock(BB: ExitBB, /*IsFinished=*/true); |
| 1424 | } |
| 1425 | |
| 1426 | /// Cast value to the specified type. |
| 1427 | static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val, |
| 1428 | QualType ValTy, QualType CastTy, |
| 1429 | SourceLocation Loc) { |
| 1430 | assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && |
| 1431 | "Cast type must sized." ); |
| 1432 | assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && |
| 1433 | "Val type must sized." ); |
| 1434 | llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(T: CastTy); |
| 1435 | if (ValTy == CastTy) |
| 1436 | return Val; |
| 1437 | if (CGF.getContext().getTypeSizeInChars(T: ValTy) == |
| 1438 | CGF.getContext().getTypeSizeInChars(T: CastTy)) |
| 1439 | return CGF.Builder.CreateBitCast(V: Val, DestTy: LLVMCastTy); |
| 1440 | if (CastTy->isIntegerType() && ValTy->isIntegerType()) |
| 1441 | return CGF.Builder.CreateIntCast(V: Val, DestTy: LLVMCastTy, |
| 1442 | isSigned: CastTy->hasSignedIntegerRepresentation()); |
| 1443 | Address CastItem = CGF.CreateMemTempWithoutCast(T: CastTy); |
| 1444 | Address ValCastItem = CastItem.withElementType(ElemTy: Val->getType()); |
| 1445 | CGF.EmitStoreOfScalar(Value: Val, Addr: ValCastItem, /*Volatile=*/false, Ty: ValTy, |
| 1446 | BaseInfo: LValueBaseInfo(AlignmentSource::Type), |
| 1447 | TBAAInfo: TBAAAccessInfo()); |
| 1448 | return CGF.EmitLoadOfScalar(Addr: CastItem, /*Volatile=*/false, Ty: CastTy, Loc, |
| 1449 | BaseInfo: LValueBaseInfo(AlignmentSource::Type), |
| 1450 | TBAAInfo: TBAAAccessInfo()); |
| 1451 | } |
| 1452 | |
| 1453 | /// Extracts the built-in reduction operator from a combiner of the form `x = x |
| 1454 | /// <op> rhs` (or the min/max conditional), or nullopt if the shape is not |
| 1455 | /// recognized (e.g. user-defined reductions). |
| 1456 | static std::optional<BinaryOperatorKind> |
| 1457 | getReductionBinOpKind(const Expr *ReductionOp) { |
| 1458 | const auto *Assign = dyn_cast<BinaryOperator>(Val: ReductionOp); |
| 1459 | if (!Assign || Assign->getOpcode() != BO_Assign) |
| 1460 | return std::nullopt; |
| 1461 | const Expr *RHS = Assign->getRHS(); |
| 1462 | // min/max are lowered as `x <cmp> rhs ? x : rhs`; the comparison identifies |
| 1463 | // it. |
| 1464 | if (const auto *ACO = |
| 1465 | dyn_cast<AbstractConditionalOperator>(Val: RHS->IgnoreParenImpCasts())) |
| 1466 | RHS = ACO->getCond(); |
| 1467 | if (const auto *BO = dyn_cast<BinaryOperator>(Val: RHS->IgnoreParenImpCasts())) |
| 1468 | return BO->getOpcode(); |
| 1469 | return std::nullopt; |
| 1470 | } |
| 1471 | |
| 1472 | /// Maps a built-in reduction operator to an atomicrmw opcode for the atomic |
| 1473 | /// cross-team reduction fast path, or nullopt if there is no direct atomicrmw |
| 1474 | /// (e.g. user-defined, complex, fp min/max) so the buffer path is used instead. |
| 1475 | static std::optional<llvm::AtomicRMWInst::BinOp> |
| 1476 | getReductionAtomicRMWOp(BinaryOperatorKind BOK, QualType Ty) { |
| 1477 | bool IsInt = Ty->isIntegerType(); |
| 1478 | bool IsSigned = Ty->hasSignedIntegerRepresentation(); |
| 1479 | switch (BOK) { |
| 1480 | case BO_Add: |
| 1481 | case BO_Sub: // A `-` reduction sums the partials, so it accumulates with add. |
| 1482 | if (IsInt) |
| 1483 | return llvm::AtomicRMWInst::Add; |
| 1484 | if (Ty->isFloatingType()) |
| 1485 | return llvm::AtomicRMWInst::FAdd; |
| 1486 | return std::nullopt; |
| 1487 | case BO_And: |
| 1488 | return IsInt ? std::optional(llvm::AtomicRMWInst::And) : std::nullopt; |
| 1489 | case BO_Or: |
| 1490 | return IsInt ? std::optional(llvm::AtomicRMWInst::Or) : std::nullopt; |
| 1491 | case BO_Xor: |
| 1492 | return IsInt ? std::optional(llvm::AtomicRMWInst::Xor) : std::nullopt; |
| 1493 | case BO_LT: // min |
| 1494 | if (IsInt) |
| 1495 | return IsSigned ? llvm::AtomicRMWInst::Min : llvm::AtomicRMWInst::UMin; |
| 1496 | return std::nullopt; |
| 1497 | case BO_GT: // max |
| 1498 | if (IsInt) |
| 1499 | return IsSigned ? llvm::AtomicRMWInst::Max : llvm::AtomicRMWInst::UMax; |
| 1500 | return std::nullopt; |
| 1501 | default: |
| 1502 | return std::nullopt; |
| 1503 | } |
| 1504 | } |
| 1505 | |
| 1506 | /// |
| 1507 | /// Design of OpenMP reductions on the GPU |
| 1508 | /// |
| 1509 | /// Consider a typical OpenMP program with one or more reduction |
| 1510 | /// clauses: |
| 1511 | /// |
| 1512 | /// float foo; |
| 1513 | /// double bar; |
| 1514 | /// #pragma omp target teams distribute parallel for \ |
| 1515 | /// reduction(+:foo) reduction(*:bar) |
| 1516 | /// for (int i = 0; i < N; i++) { |
| 1517 | /// foo += A[i]; bar *= B[i]; |
| 1518 | /// } |
| 1519 | /// |
| 1520 | /// where 'foo' and 'bar' are reduced across all OpenMP threads in |
| 1521 | /// all teams. In our OpenMP implementation on the NVPTX device an |
| 1522 | /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads |
| 1523 | /// within a team are mapped to CUDA threads within a threadblock. |
| 1524 | /// Our goal is to efficiently aggregate values across all OpenMP |
| 1525 | /// threads such that: |
| 1526 | /// |
| 1527 | /// - the compiler and runtime are logically concise, and |
| 1528 | /// - the reduction is performed efficiently in a hierarchical |
| 1529 | /// manner as follows: within OpenMP threads in the same warp, |
| 1530 | /// across warps in a threadblock, and finally across teams on |
| 1531 | /// the NVPTX device. |
| 1532 | /// |
| 1533 | /// Introduction to Decoupling |
| 1534 | /// |
| 1535 | /// We would like to decouple the compiler and the runtime so that the |
| 1536 | /// latter is ignorant of the reduction variables (number, data types) |
| 1537 | /// and the reduction operators. This allows a simpler interface |
| 1538 | /// and implementation while still attaining good performance. |
| 1539 | /// |
| 1540 | /// Pseudocode for the aforementioned OpenMP program generated by the |
| 1541 | /// compiler is as follows: |
| 1542 | /// |
| 1543 | /// 1. Create private copies of reduction variables on each OpenMP |
| 1544 | /// thread: 'foo_private', 'bar_private' |
| 1545 | /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned |
| 1546 | /// to it and writes the result in 'foo_private' and 'bar_private' |
| 1547 | /// respectively. |
| 1548 | /// 3. Call the OpenMP runtime on the GPU to reduce within a team |
| 1549 | /// and store the result on the team master: |
| 1550 | /// |
| 1551 | /// __kmpc_nvptx_parallel_reduce_nowait_v2(..., |
| 1552 | /// reduceData, shuffleReduceFn, interWarpCpyFn) |
| 1553 | /// |
| 1554 | /// where: |
| 1555 | /// struct ReduceData { |
| 1556 | /// double *foo; |
| 1557 | /// double *bar; |
| 1558 | /// } reduceData |
| 1559 | /// reduceData.foo = &foo_private |
| 1560 | /// reduceData.bar = &bar_private |
| 1561 | /// |
| 1562 | /// 'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two |
| 1563 | /// auxiliary functions generated by the compiler that operate on |
| 1564 | /// variables of type 'ReduceData'. They aid the runtime perform |
| 1565 | /// algorithmic steps in a data agnostic manner. |
| 1566 | /// |
| 1567 | /// 'shuffleReduceFn' is a pointer to a function that reduces data |
| 1568 | /// of type 'ReduceData' across two OpenMP threads (lanes) in the |
| 1569 | /// same warp. It takes the following arguments as input: |
| 1570 | /// |
| 1571 | /// a. variable of type 'ReduceData' on the calling lane, |
| 1572 | /// b. its lane_id, |
| 1573 | /// c. an offset relative to the current lane_id to generate a |
| 1574 | /// remote_lane_id. The remote lane contains the second |
| 1575 | /// variable of type 'ReduceData' that is to be reduced. |
| 1576 | /// d. an algorithm version parameter determining which reduction |
| 1577 | /// algorithm to use. |
| 1578 | /// |
| 1579 | /// 'shuffleReduceFn' retrieves data from the remote lane using |
| 1580 | /// efficient GPU shuffle intrinsics and reduces, using the |
| 1581 | /// algorithm specified by the 4th parameter, the two operands |
| 1582 | /// element-wise. The result is written to the first operand. |
| 1583 | /// |
| 1584 | /// Different reduction algorithms are implemented in different |
| 1585 | /// runtime functions, all calling 'shuffleReduceFn' to perform |
| 1586 | /// the essential reduction step. Therefore, based on the 4th |
| 1587 | /// parameter, this function behaves slightly differently to |
| 1588 | /// cooperate with the runtime to ensure correctness under |
| 1589 | /// different circumstances. |
| 1590 | /// |
| 1591 | /// 'InterWarpCpyFn' is a pointer to a function that transfers |
| 1592 | /// reduced variables across warps. It tunnels, through CUDA |
| 1593 | /// shared memory, the thread-private data of type 'ReduceData' |
| 1594 | /// from lane 0 of each warp to a lane in the first warp. |
| 1595 | /// 4. Call the OpenMP runtime on the GPU to reduce across teams. |
| 1596 | /// The last team writes the global reduced value to memory. |
| 1597 | /// |
| 1598 | /// ret = __kmpc_nvptx_teams_reduce_nowait(..., |
| 1599 | /// reduceData, shuffleReduceFn, interWarpCpyFn, |
| 1600 | /// scratchpadCopyFn, loadAndReduceFn) |
| 1601 | /// |
| 1602 | /// 'scratchpadCopyFn' is a helper that stores reduced |
| 1603 | /// data from the team master to a scratchpad array in |
| 1604 | /// global memory. |
| 1605 | /// |
| 1606 | /// 'loadAndReduceFn' is a helper that loads data from |
| 1607 | /// the scratchpad array and reduces it with the input |
| 1608 | /// operand. |
| 1609 | /// |
| 1610 | /// These compiler generated functions hide address |
| 1611 | /// calculation and alignment information from the runtime. |
| 1612 | /// 5. if ret == 1: |
| 1613 | /// The team master of the last team stores the reduced |
| 1614 | /// result to the globals in memory. |
| 1615 | /// foo += reduceData.foo; bar *= reduceData.bar |
| 1616 | /// |
| 1617 | /// |
| 1618 | /// Warp Reduction Algorithms |
| 1619 | /// |
| 1620 | /// On the warp level, we have three algorithms implemented in the |
| 1621 | /// OpenMP runtime depending on the number of active lanes: |
| 1622 | /// |
| 1623 | /// Full Warp Reduction |
| 1624 | /// |
| 1625 | /// The reduce algorithm within a warp where all lanes are active |
| 1626 | /// is implemented in the runtime as follows: |
| 1627 | /// |
| 1628 | /// full_warp_reduce(void *reduce_data, |
| 1629 | /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { |
| 1630 | /// for (int offset = WARPSIZE/2; offset > 0; offset /= 2) |
| 1631 | /// ShuffleReduceFn(reduce_data, 0, offset, 0); |
| 1632 | /// } |
| 1633 | /// |
| 1634 | /// The algorithm completes in log(2, WARPSIZE) steps. |
| 1635 | /// |
| 1636 | /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is |
| 1637 | /// not used therefore we save instructions by not retrieving lane_id |
| 1638 | /// from the corresponding special registers. The 4th parameter, which |
| 1639 | /// represents the version of the algorithm being used, is set to 0 to |
| 1640 | /// signify full warp reduction. |
| 1641 | /// |
| 1642 | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
| 1643 | /// |
| 1644 | /// #reduce_elem refers to an element in the local lane's data structure |
| 1645 | /// #remote_elem is retrieved from a remote lane |
| 1646 | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
| 1647 | /// reduce_elem = reduce_elem REDUCE_OP remote_elem; |
| 1648 | /// |
| 1649 | /// Contiguous Partial Warp Reduction |
| 1650 | /// |
| 1651 | /// This reduce algorithm is used within a warp where only the first |
| 1652 | /// 'n' (n <= WARPSIZE) lanes are active. It is typically used when the |
| 1653 | /// number of OpenMP threads in a parallel region is not a multiple of |
| 1654 | /// WARPSIZE. The algorithm is implemented in the runtime as follows: |
| 1655 | /// |
| 1656 | /// void |
| 1657 | /// contiguous_partial_reduce(void *reduce_data, |
| 1658 | /// kmp_ShuffleReductFctPtr ShuffleReduceFn, |
| 1659 | /// int size, int lane_id) { |
| 1660 | /// int curr_size; |
| 1661 | /// int offset; |
| 1662 | /// curr_size = size; |
| 1663 | /// mask = curr_size/2; |
| 1664 | /// while (offset>0) { |
| 1665 | /// ShuffleReduceFn(reduce_data, lane_id, offset, 1); |
| 1666 | /// curr_size = (curr_size+1)/2; |
| 1667 | /// offset = curr_size/2; |
| 1668 | /// } |
| 1669 | /// } |
| 1670 | /// |
| 1671 | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
| 1672 | /// |
| 1673 | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
| 1674 | /// if (lane_id < offset) |
| 1675 | /// reduce_elem = reduce_elem REDUCE_OP remote_elem |
| 1676 | /// else |
| 1677 | /// reduce_elem = remote_elem |
| 1678 | /// |
| 1679 | /// This algorithm assumes that the data to be reduced are located in a |
| 1680 | /// contiguous subset of lanes starting from the first. When there is |
| 1681 | /// an odd number of active lanes, the data in the last lane is not |
| 1682 | /// aggregated with any other lane's dat but is instead copied over. |
| 1683 | /// |
| 1684 | /// Dispersed Partial Warp Reduction |
| 1685 | /// |
| 1686 | /// This algorithm is used within a warp when any discontiguous subset of |
| 1687 | /// lanes are active. It is used to implement the reduction operation |
| 1688 | /// across lanes in an OpenMP simd region or in a nested parallel region. |
| 1689 | /// |
| 1690 | /// void |
| 1691 | /// dispersed_partial_reduce(void *reduce_data, |
| 1692 | /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { |
| 1693 | /// int size, remote_id; |
| 1694 | /// int logical_lane_id = number_of_active_lanes_before_me() * 2; |
| 1695 | /// do { |
| 1696 | /// remote_id = next_active_lane_id_right_after_me(); |
| 1697 | /// # the above function returns 0 of no active lane |
| 1698 | /// # is present right after the current lane. |
| 1699 | /// size = number_of_active_lanes_in_this_warp(); |
| 1700 | /// logical_lane_id /= 2; |
| 1701 | /// ShuffleReduceFn(reduce_data, logical_lane_id, |
| 1702 | /// remote_id-1-threadIdx.x, 2); |
| 1703 | /// } while (logical_lane_id % 2 == 0 && size > 1); |
| 1704 | /// } |
| 1705 | /// |
| 1706 | /// There is no assumption made about the initial state of the reduction. |
| 1707 | /// Any number of lanes (>=1) could be active at any position. The reduction |
| 1708 | /// result is returned in the first active lane. |
| 1709 | /// |
| 1710 | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
| 1711 | /// |
| 1712 | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
| 1713 | /// if (lane_id % 2 == 0 && offset > 0) |
| 1714 | /// reduce_elem = reduce_elem REDUCE_OP remote_elem |
| 1715 | /// else |
| 1716 | /// reduce_elem = remote_elem |
| 1717 | /// |
| 1718 | /// |
| 1719 | /// Intra-Team Reduction |
| 1720 | /// |
| 1721 | /// This function, as implemented in the runtime call |
| 1722 | /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP |
| 1723 | /// threads in a team. It first reduces within a warp using the |
| 1724 | /// aforementioned algorithms. We then proceed to gather all such |
| 1725 | /// reduced values at the first warp. |
| 1726 | /// |
| 1727 | /// The runtime makes use of the function 'InterWarpCpyFn', which copies |
| 1728 | /// data from each of the "warp master" (zeroth lane of each warp, where |
| 1729 | /// warp-reduced data is held) to the zeroth warp. This step reduces (in |
| 1730 | /// a mathematical sense) the problem of reduction across warp masters in |
| 1731 | /// a block to the problem of warp reduction. |
| 1732 | /// |
| 1733 | /// |
| 1734 | /// Inter-Team Reduction |
| 1735 | /// |
| 1736 | /// Once a team has reduced its data to a single value, it is stored in |
| 1737 | /// a global scratchpad array. Since each team has a distinct slot, this |
| 1738 | /// can be done without locking. |
| 1739 | /// |
| 1740 | /// The last team to write to the scratchpad array proceeds to reduce the |
| 1741 | /// scratchpad array. One or more workers in the last team use the helper |
| 1742 | /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e., |
| 1743 | /// the k'th worker reduces every k'th element. |
| 1744 | /// |
| 1745 | /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to |
| 1746 | /// reduce across workers and compute a globally reduced value. |
| 1747 | /// |
| 1748 | void CGOpenMPRuntimeGPU::emitReduction( |
| 1749 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, |
| 1750 | ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, |
| 1751 | ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { |
| 1752 | if (!CGF.HaveInsertPoint()) |
| 1753 | return; |
| 1754 | |
| 1755 | bool ParallelReduction = isOpenMPParallelDirective(DKind: Options.ReductionKind); |
| 1756 | bool TeamsReduction = isOpenMPTeamsDirective(DKind: Options.ReductionKind); |
| 1757 | |
| 1758 | if (Options.SimpleReduction) { |
| 1759 | assert(!TeamsReduction && !ParallelReduction && |
| 1760 | "Invalid reduction selection in emitReduction." ); |
| 1761 | (void)ParallelReduction; |
| 1762 | CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, |
| 1763 | ReductionOps, Options); |
| 1764 | return; |
| 1765 | } |
| 1766 | |
| 1767 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap; |
| 1768 | llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size()); |
| 1769 | int Cnt = 0; |
| 1770 | for (const Expr *DRE : Privates) { |
| 1771 | PrivatesReductions[Cnt] = cast<DeclRefExpr>(Val: DRE)->getDecl(); |
| 1772 | ++Cnt; |
| 1773 | } |
| 1774 | const RecordDecl *ReductionRec = ::buildRecordForGlobalizedVars( |
| 1775 | C&: CGM.getContext(), EscapedDecls: PrivatesReductions, EscapedDeclsForTeams: {}, MappedDeclsFields&: VarFieldMap, BufSize: 1); |
| 1776 | |
| 1777 | // The atomic cross-team reduction fast path is opt-in. Hand each eligible |
| 1778 | // scalar reduction an atomic combiner; createReductionsGPU uses the atomic |
| 1779 | // path only if every reduction in the set has one. Track whether that holds |
| 1780 | // so we can skip the (then unused) per-team buffer registration. |
| 1781 | bool UseAtomicReduction = |
| 1782 | TeamsReduction && CGM.getLangOpts().OpenMPTargetAtomicReduction; |
| 1783 | bool AllAtomicable = UseAtomicReduction; |
| 1784 | |
| 1785 | // Source location for the ident struct |
| 1786 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); |
| 1787 | |
| 1788 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 1789 | InsertPointTy AllocaIP(CGF.AllocaInsertPt->getParent(), |
| 1790 | CGF.AllocaInsertPt->getIterator()); |
| 1791 | InsertPointTy CodeGenIP(CGF.Builder.GetInsertBlock(), |
| 1792 | CGF.Builder.GetInsertPoint()); |
| 1793 | llvm::OpenMPIRBuilder::LocationDescription OmpLoc( |
| 1794 | CodeGenIP, CGF.SourceLocToDebugLoc(Location: Loc)); |
| 1795 | llvm::SmallVector<llvm::OpenMPIRBuilder::ReductionInfo, 2> ReductionInfos; |
| 1796 | |
| 1797 | CodeGenFunction::OMPPrivateScope Scope(CGF); |
| 1798 | unsigned Idx = 0; |
| 1799 | for (const Expr *Private : Privates) { |
| 1800 | llvm::Type *ElementType; |
| 1801 | llvm::Value *Variable; |
| 1802 | llvm::Value *PrivateVariable; |
| 1803 | llvm::OpenMPIRBuilder::ReductionGenAtomicCBTy AtomicReductionGen = nullptr; |
| 1804 | ElementType = CGF.ConvertTypeForMem(T: Private->getType()); |
| 1805 | const auto *RHSVar = |
| 1806 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: RHSExprs[Idx])->getDecl()); |
| 1807 | PrivateVariable = CGF.GetAddrOfLocalVar(VD: RHSVar).emitRawPointer(CGF); |
| 1808 | const auto *LHSVar = |
| 1809 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: LHSExprs[Idx])->getDecl()); |
| 1810 | Variable = CGF.GetAddrOfLocalVar(VD: LHSVar).emitRawPointer(CGF); |
| 1811 | llvm::OpenMPIRBuilder::EvalKind EvalKind; |
| 1812 | switch (CGF.getEvaluationKind(T: Private->getType())) { |
| 1813 | case TEK_Scalar: |
| 1814 | EvalKind = llvm::OpenMPIRBuilder::EvalKind::Scalar; |
| 1815 | break; |
| 1816 | case TEK_Complex: |
| 1817 | EvalKind = llvm::OpenMPIRBuilder::EvalKind::Complex; |
| 1818 | break; |
| 1819 | case TEK_Aggregate: |
| 1820 | EvalKind = llvm::OpenMPIRBuilder::EvalKind::Aggregate; |
| 1821 | break; |
| 1822 | } |
| 1823 | auto ReductionGen = [&](InsertPointTy CodeGenIP, unsigned I, |
| 1824 | llvm::Value **LHSPtr, llvm::Value **RHSPtr, |
| 1825 | llvm::Function *NewFunc) { |
| 1826 | CGF.Builder.restoreIP(IP: CodeGenIP); |
| 1827 | auto *CurFn = CGF.CurFn; |
| 1828 | CGF.CurFn = NewFunc; |
| 1829 | |
| 1830 | // The helper has no DISubprogram of its own, so a debug location here |
| 1831 | // would name the enclosing function's scope, which is invalid IR. |
| 1832 | // Suppress them, as the other OpenMPIRBuilder-generated helpers do. |
| 1833 | llvm::DebugLoc SavedDebugLoc = CGF.Builder.getCurrentDebugLocation(); |
| 1834 | CGF.Builder.SetCurrentDebugLocation(llvm::DebugLoc()); |
| 1835 | CGF.disableDebugInfo(); |
| 1836 | |
| 1837 | *LHSPtr = CGF.GetAddrOfLocalVar( |
| 1838 | VD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: LHSExprs[I])->getDecl())) |
| 1839 | .emitRawPointer(CGF); |
| 1840 | *RHSPtr = CGF.GetAddrOfLocalVar( |
| 1841 | VD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: RHSExprs[I])->getDecl())) |
| 1842 | .emitRawPointer(CGF); |
| 1843 | |
| 1844 | emitSingleReductionCombiner(CGF, ReductionOp: ReductionOps[I], PrivateRef: Privates[I], |
| 1845 | LHS: cast<DeclRefExpr>(Val: LHSExprs[I]), |
| 1846 | RHS: cast<DeclRefExpr>(Val: RHSExprs[I])); |
| 1847 | |
| 1848 | CGF.enableDebugInfo(); |
| 1849 | CGF.Builder.SetCurrentDebugLocation(SavedDebugLoc); |
| 1850 | CGF.CurFn = CurFn; |
| 1851 | |
| 1852 | return InsertPointTy(CGF.Builder.GetInsertBlock(), |
| 1853 | CGF.Builder.GetInsertPoint()); |
| 1854 | }; |
| 1855 | |
| 1856 | // For the atomic fast path, hand this reduction an atomic combiner if it is |
| 1857 | // a scalar with a direct atomicrmw; otherwise the set is not fully |
| 1858 | // atomicable and falls back to the buffer path. |
| 1859 | if (UseAtomicReduction) { |
| 1860 | std::optional<llvm::AtomicRMWInst::BinOp> AtomicOp; |
| 1861 | if (EvalKind == llvm::OpenMPIRBuilder::EvalKind::Scalar) { |
| 1862 | if (std::optional<BinaryOperatorKind> BOK = |
| 1863 | getReductionBinOpKind(ReductionOp: ReductionOps[Idx])) |
| 1864 | AtomicOp = getReductionAtomicRMWOp(BOK: *BOK, Ty: Private->getType()); |
| 1865 | } |
| 1866 | if (!AtomicOp) { |
| 1867 | AllAtomicable = false; |
| 1868 | } else { |
| 1869 | llvm::AtomicRMWInst::BinOp Op = *AtomicOp; |
| 1870 | llvm::Align Alignment = |
| 1871 | CGM.getModule().getDataLayout().getPrefTypeAlign(Ty: ElementType); |
| 1872 | // Device (agent) scope suffices: all teams accumulate on-device and the |
| 1873 | // host reads the result only after the kernel (via map-back), so the |
| 1874 | // far costlier system scope is unnecessary. The |
| 1875 | // no.fine.grained/no.remote memory metadata is omitted so the atomic |
| 1876 | // stays correct under USM. |
| 1877 | llvm::SyncScope::ID SSID = CGF.getTargetHooks().getLLVMSyncScopeID( |
| 1878 | LangOpts: CGF.getLangOpts(), Scope: SyncScope::DeviceScope, |
| 1879 | Ordering: llvm::AtomicOrdering::Monotonic, Ctx&: CGF.getLLVMContext()); |
| 1880 | AtomicReductionGen = [Op, Alignment, |
| 1881 | SSID](InsertPointTy IP, llvm::Type *EltTy, |
| 1882 | llvm::Value *LHS, llvm::Value *RHS) |
| 1883 | -> llvm::OpenMPIRBuilder::InsertPointOrErrorTy { |
| 1884 | llvm::IRBuilder<> Builder(IP.getBlock(), IP.getPoint()); |
| 1885 | llvm::Value *Val = Builder.CreateLoad(Ty: EltTy, Ptr: RHS); |
| 1886 | Builder.CreateAtomicRMW(Op, Ptr: LHS, Val, Align: Alignment, |
| 1887 | Ordering: llvm::AtomicOrdering::Monotonic, SSID); |
| 1888 | return InsertPointTy(Builder.GetInsertBlock(), |
| 1889 | Builder.GetInsertPoint()); |
| 1890 | }; |
| 1891 | } |
| 1892 | } |
| 1893 | |
| 1894 | ReductionInfos.emplace_back(Args: llvm::OpenMPIRBuilder::ReductionInfo( |
| 1895 | ElementType, Variable, PrivateVariable, EvalKind, |
| 1896 | /*ReductionGen=*/nullptr, ReductionGen, AtomicReductionGen, |
| 1897 | /*DataPtrPtrGen=*/nullptr)); |
| 1898 | Idx++; |
| 1899 | } |
| 1900 | |
| 1901 | // The atomic path folds directly into the mapped variable and needs no |
| 1902 | // per-team buffer; register the record for buffer allocation otherwise. |
| 1903 | if (TeamsReduction && !AllAtomicable) |
| 1904 | TeamsReductions.push_back(Elt: ReductionRec); |
| 1905 | |
| 1906 | bool IsSPMD = getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD; |
| 1907 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = |
| 1908 | cantFail(ValOrErr: OMPBuilder.createReductionsGPU( |
| 1909 | Loc: OmpLoc, AllocaIP, CodeGenIP, ReductionInfos, /*IsByRef=*/{}, IsNoWait: false, |
| 1910 | IsTeamsReduction: TeamsReduction, IsSPMD, |
| 1911 | ReductionGenCBKind: llvm::OpenMPIRBuilder::ReductionGenCBKind::Clang, |
| 1912 | GridValue: CGF.getTarget().getGridValue(), SrcLocInfo: RTLoc)); |
| 1913 | CGF.Builder.restoreIP(IP: AfterIP); |
| 1914 | } |
| 1915 | |
| 1916 | const VarDecl * |
| 1917 | CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD, |
| 1918 | const VarDecl *NativeParam) const { |
| 1919 | if (!NativeParam->getType()->isReferenceType()) |
| 1920 | return NativeParam; |
| 1921 | QualType ArgType = NativeParam->getType(); |
| 1922 | QualifierCollector QC; |
| 1923 | const Type *NonQualTy = QC.strip(type: ArgType); |
| 1924 | QualType PointeeTy = cast<ReferenceType>(Val: NonQualTy)->getPointeeType(); |
| 1925 | if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) { |
| 1926 | if (Attr->getCaptureKind() == OMPC_map) { |
| 1927 | PointeeTy = CGM.getContext().getAddrSpaceQualType(T: PointeeTy, |
| 1928 | AddressSpace: LangAS::opencl_global); |
| 1929 | } |
| 1930 | } |
| 1931 | ArgType = CGM.getContext().getPointerType(T: PointeeTy); |
| 1932 | QC.addRestrict(); |
| 1933 | ArgType = QC.apply(Context: CGM.getContext(), QT: ArgType); |
| 1934 | if (isa<ImplicitParamDecl>(Val: NativeParam)) |
| 1935 | return ImplicitParamDecl::Create( |
| 1936 | C&: CGM.getContext(), /*DC=*/nullptr, IdLoc: NativeParam->getLocation(), |
| 1937 | Id: NativeParam->getIdentifier(), T: ArgType, ParamKind: ImplicitParamKind::Other); |
| 1938 | return ParmVarDecl::Create( |
| 1939 | C&: CGM.getContext(), |
| 1940 | DC: const_cast<DeclContext *>(NativeParam->getDeclContext()), |
| 1941 | StartLoc: NativeParam->getBeginLoc(), IdLoc: NativeParam->getLocation(), |
| 1942 | Id: NativeParam->getIdentifier(), T: ArgType, |
| 1943 | /*TInfo=*/nullptr, S: SC_None, /*DefArg=*/nullptr); |
| 1944 | } |
| 1945 | |
| 1946 | Address |
| 1947 | CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF, |
| 1948 | const VarDecl *NativeParam, |
| 1949 | const VarDecl *TargetParam) const { |
| 1950 | assert(NativeParam != TargetParam && |
| 1951 | NativeParam->getType()->isReferenceType() && |
| 1952 | "Native arg must not be the same as target arg." ); |
| 1953 | Address LocalAddr = CGF.GetAddrOfLocalVar(VD: TargetParam); |
| 1954 | QualType NativeParamType = NativeParam->getType(); |
| 1955 | QualifierCollector QC; |
| 1956 | const Type *NonQualTy = QC.strip(type: NativeParamType); |
| 1957 | QualType NativePointeeTy = cast<ReferenceType>(Val: NonQualTy)->getPointeeType(); |
| 1958 | unsigned NativePointeeAddrSpace = |
| 1959 | CGF.getTypes().getTargetAddressSpace(T: NativePointeeTy); |
| 1960 | QualType TargetTy = TargetParam->getType(); |
| 1961 | llvm::Value *TargetAddr = CGF.EmitLoadOfScalar(Addr: LocalAddr, /*Volatile=*/false, |
| 1962 | Ty: TargetTy, Loc: SourceLocation()); |
| 1963 | // Cast to native address space. |
| 1964 | TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 1965 | V: TargetAddr, |
| 1966 | DestTy: llvm::PointerType::get(C&: CGF.getLLVMContext(), AddressSpace: NativePointeeAddrSpace)); |
| 1967 | Address NativeParamAddr = CGF.CreateMemTemp(T: NativeParamType); |
| 1968 | CGF.EmitStoreOfScalar(Value: TargetAddr, Addr: NativeParamAddr, /*Volatile=*/false, |
| 1969 | Ty: NativeParamType); |
| 1970 | return NativeParamAddr; |
| 1971 | } |
| 1972 | |
| 1973 | void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall( |
| 1974 | CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, |
| 1975 | ArrayRef<llvm::Value *> Args) const { |
| 1976 | SmallVector<llvm::Value *, 4> TargetArgs; |
| 1977 | TargetArgs.reserve(N: Args.size()); |
| 1978 | auto *FnType = OutlinedFn.getFunctionType(); |
| 1979 | for (unsigned I = 0, E = Args.size(); I < E; ++I) { |
| 1980 | if (FnType->isVarArg() && FnType->getNumParams() <= I) { |
| 1981 | TargetArgs.append(in_start: std::next(x: Args.begin(), n: I), in_end: Args.end()); |
| 1982 | break; |
| 1983 | } |
| 1984 | llvm::Type *TargetType = FnType->getParamType(i: I); |
| 1985 | llvm::Value *NativeArg = Args[I]; |
| 1986 | if (!TargetType->isPointerTy()) { |
| 1987 | TargetArgs.emplace_back(Args&: NativeArg); |
| 1988 | continue; |
| 1989 | } |
| 1990 | TargetArgs.emplace_back( |
| 1991 | Args: CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(V: NativeArg, DestTy: TargetType)); |
| 1992 | } |
| 1993 | CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, Args: TargetArgs); |
| 1994 | } |
| 1995 | |
| 1996 | /// Emit function which wraps the outline parallel region |
| 1997 | /// and controls the arguments which are passed to this function. |
| 1998 | /// The wrapper ensures that the outlined function is called |
| 1999 | /// with the correct arguments when data is shared. |
| 2000 | llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper( |
| 2001 | llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) { |
| 2002 | ASTContext &Ctx = CGM.getContext(); |
| 2003 | const auto &CS = *D.getCapturedStmt(RegionKind: OMPD_parallel); |
| 2004 | |
| 2005 | // Create a function that takes as argument the source thread. |
| 2006 | FunctionArgList WrapperArgs; |
| 2007 | QualType Int16QTy = |
| 2008 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false); |
| 2009 | QualType Int32QTy = |
| 2010 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false); |
| 2011 | auto *ParallelLevelArg = ImplicitParamDecl::Create( |
| 2012 | C&: Ctx, /*DC=*/nullptr, IdLoc: D.getBeginLoc(), |
| 2013 | /*Id=*/nullptr, T: Int16QTy, ParamKind: ImplicitParamKind::Other); |
| 2014 | auto *WrapperArg = ImplicitParamDecl::Create( |
| 2015 | C&: Ctx, /*DC=*/nullptr, IdLoc: D.getBeginLoc(), |
| 2016 | /*Id=*/nullptr, T: Int32QTy, ParamKind: ImplicitParamKind::Other); |
| 2017 | WrapperArgs.emplace_back(Args&: ParallelLevelArg); |
| 2018 | WrapperArgs.emplace_back(Args&: WrapperArg); |
| 2019 | |
| 2020 | const CGFunctionInfo &CGFI = |
| 2021 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(resultType: Ctx.VoidTy, args: WrapperArgs); |
| 2022 | |
| 2023 | auto *Fn = llvm::Function::Create( |
| 2024 | Ty: CGM.getTypes().GetFunctionType(Info: CGFI), Linkage: llvm::GlobalValue::InternalLinkage, |
| 2025 | N: Twine(OutlinedParallelFn->getName(), "_wrapper" ), M: &CGM.getModule()); |
| 2026 | |
| 2027 | // Ensure we do not inline the function. This is trivially true for the ones |
| 2028 | // passed to __kmpc_fork_call but the ones calles in serialized regions |
| 2029 | // could be inlined. This is not a perfect but it is closer to the invariant |
| 2030 | // we want, namely, every data environment starts with a new function. |
| 2031 | // TODO: We should pass the if condition to the runtime function and do the |
| 2032 | // handling there. Much cleaner code. |
| 2033 | Fn->addFnAttr(Kind: llvm::Attribute::NoInline); |
| 2034 | |
| 2035 | CGM.SetInternalFunctionAttributes(GD: GlobalDecl(), F: Fn, FI: CGFI); |
| 2036 | Fn->setLinkage(llvm::GlobalValue::InternalLinkage); |
| 2037 | Fn->setDoesNotRecurse(); |
| 2038 | |
| 2039 | CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); |
| 2040 | CGF.StartFunction(GD: GlobalDecl(), RetTy: Ctx.VoidTy, Fn, FnInfo: CGFI, Args: WrapperArgs, |
| 2041 | Loc: D.getBeginLoc(), StartLoc: D.getBeginLoc()); |
| 2042 | |
| 2043 | const auto *RD = CS.getCapturedRecordDecl(); |
| 2044 | auto CurField = RD->field_begin(); |
| 2045 | |
| 2046 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(Ty: CGF.Int32Ty, |
| 2047 | /*Name=*/".zero.addr" ); |
| 2048 | CGF.Builder.CreateStore(Val: CGF.Builder.getInt32(/*C*/ 0), Addr: ZeroAddr); |
| 2049 | // Get the array of arguments. |
| 2050 | SmallVector<llvm::Value *, 8> Args; |
| 2051 | |
| 2052 | Args.emplace_back(Args: CGF.GetAddrOfLocalVar(VD: WrapperArg).emitRawPointer(CGF)); |
| 2053 | Args.emplace_back(Args: ZeroAddr.emitRawPointer(CGF)); |
| 2054 | |
| 2055 | CGBuilderTy &Bld = CGF.Builder; |
| 2056 | auto CI = CS.capture_begin(); |
| 2057 | |
| 2058 | // Use global memory for data sharing. |
| 2059 | // Handle passing of global args to workers. |
| 2060 | RawAddress GlobalArgs = |
| 2061 | CGF.CreateDefaultAlignTempAlloca(Ty: CGF.VoidPtrPtrTy, Name: "global_args" ); |
| 2062 | llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer(); |
| 2063 | llvm::Value *DataSharingArgs[] = {GlobalArgsPtr}; |
| 2064 | CGF.EmitRuntimeCall(callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 2065 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_get_shared_variables), |
| 2066 | args: DataSharingArgs); |
| 2067 | |
| 2068 | // Retrieve the shared variables from the list of references returned |
| 2069 | // by the runtime. Pass the variables to the outlined function. |
| 2070 | Address SharedArgListAddress = Address::invalid(); |
| 2071 | if (CS.capture_size() > 0 || |
| 2072 | isOpenMPLoopBoundSharingDirective(Kind: D.getDirectiveKind())) { |
| 2073 | SharedArgListAddress = CGF.EmitLoadOfPointer( |
| 2074 | Ptr: GlobalArgs, PtrTy: CGF.getContext() |
| 2075 | .getPointerType(T: CGF.getContext().VoidPtrTy) |
| 2076 | .castAs<PointerType>()); |
| 2077 | } |
| 2078 | unsigned Idx = 0; |
| 2079 | if (isOpenMPLoopBoundSharingDirective(Kind: D.getDirectiveKind())) { |
| 2080 | Address Src = Bld.CreateConstInBoundsGEP(Addr: SharedArgListAddress, Index: Idx); |
| 2081 | Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2082 | Addr: Src, Ty: Bld.getPtrTy(AddrSpace: 0), ElementTy: CGF.SizeTy); |
| 2083 | llvm::Value *LB = CGF.EmitLoadOfScalar( |
| 2084 | Addr: TypedAddress, |
| 2085 | /*Volatile=*/false, |
| 2086 | Ty: CGF.getContext().getPointerType(T: CGF.getContext().getSizeType()), |
| 2087 | Loc: cast<OMPLoopDirective>(Val: D).getLowerBoundVariable()->getExprLoc()); |
| 2088 | Args.emplace_back(Args&: LB); |
| 2089 | ++Idx; |
| 2090 | Src = Bld.CreateConstInBoundsGEP(Addr: SharedArgListAddress, Index: Idx); |
| 2091 | TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast(Addr: Src, Ty: Bld.getPtrTy(AddrSpace: 0), |
| 2092 | ElementTy: CGF.SizeTy); |
| 2093 | llvm::Value *UB = CGF.EmitLoadOfScalar( |
| 2094 | Addr: TypedAddress, |
| 2095 | /*Volatile=*/false, |
| 2096 | Ty: CGF.getContext().getPointerType(T: CGF.getContext().getSizeType()), |
| 2097 | Loc: cast<OMPLoopDirective>(Val: D).getUpperBoundVariable()->getExprLoc()); |
| 2098 | Args.emplace_back(Args&: UB); |
| 2099 | ++Idx; |
| 2100 | } |
| 2101 | if (CS.capture_size() > 0) { |
| 2102 | ASTContext &CGFContext = CGF.getContext(); |
| 2103 | for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) { |
| 2104 | QualType ElemTy = CurField->getType(); |
| 2105 | Address Src = Bld.CreateConstInBoundsGEP(Addr: SharedArgListAddress, Index: I + Idx); |
| 2106 | Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2107 | Addr: Src, Ty: CGF.ConvertTypeForMem(T: CGFContext.getPointerType(T: ElemTy)), |
| 2108 | ElementTy: CGF.ConvertTypeForMem(T: ElemTy)); |
| 2109 | llvm::Value *Arg = CGF.EmitLoadOfScalar(Addr: TypedAddress, |
| 2110 | /*Volatile=*/false, |
| 2111 | Ty: CGFContext.getPointerType(T: ElemTy), |
| 2112 | Loc: CI->getLocation()); |
| 2113 | if (CI->capturesVariableByCopy() && |
| 2114 | !CI->getCapturedVar()->getType()->isAnyPointerType()) { |
| 2115 | Arg = castValueToType(CGF, Val: Arg, ValTy: ElemTy, CastTy: CGFContext.getUIntPtrType(), |
| 2116 | Loc: CI->getLocation()); |
| 2117 | } |
| 2118 | Args.emplace_back(Args&: Arg); |
| 2119 | } |
| 2120 | } |
| 2121 | |
| 2122 | emitOutlinedFunctionCall(CGF, Loc: D.getBeginLoc(), OutlinedFn: OutlinedParallelFn, Args); |
| 2123 | CGF.FinishFunction(); |
| 2124 | return Fn; |
| 2125 | } |
| 2126 | |
| 2127 | void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF, |
| 2128 | const Decl *D) { |
| 2129 | if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic) |
| 2130 | return; |
| 2131 | |
| 2132 | assert(D && "Expected function or captured|block decl." ); |
| 2133 | assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && |
| 2134 | "Function is registered already." ); |
| 2135 | assert((!TeamAndReductions.first || TeamAndReductions.first == D) && |
| 2136 | "Team is set but not processed." ); |
| 2137 | const Stmt *Body = nullptr; |
| 2138 | bool NeedToDelayGlobalization = false; |
| 2139 | if (const auto *FD = dyn_cast<FunctionDecl>(Val: D)) { |
| 2140 | Body = FD->getBody(); |
| 2141 | } else if (const auto *BD = dyn_cast<BlockDecl>(Val: D)) { |
| 2142 | Body = BD->getBody(); |
| 2143 | } else if (const auto *CD = dyn_cast<CapturedDecl>(Val: D)) { |
| 2144 | Body = CD->getBody(); |
| 2145 | NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP; |
| 2146 | if (NeedToDelayGlobalization && |
| 2147 | getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) |
| 2148 | return; |
| 2149 | } |
| 2150 | if (!Body) |
| 2151 | return; |
| 2152 | CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second); |
| 2153 | VarChecker.Visit(S: Body); |
| 2154 | const RecordDecl *GlobalizedVarsRecord = |
| 2155 | VarChecker.getGlobalizedRecord(IsInTTDRegion); |
| 2156 | TeamAndReductions.first = nullptr; |
| 2157 | TeamAndReductions.second.clear(); |
| 2158 | ArrayRef<const ValueDecl *> EscapedVariableLengthDecls = |
| 2159 | VarChecker.getEscapedVariableLengthDecls(); |
| 2160 | ArrayRef<const ValueDecl *> DelayedVariableLengthDecls = |
| 2161 | VarChecker.getDelayedVariableLengthDecls(); |
| 2162 | if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty() && |
| 2163 | DelayedVariableLengthDecls.empty()) |
| 2164 | return; |
| 2165 | auto I = FunctionGlobalizedDecls.try_emplace(Key: CGF.CurFn).first; |
| 2166 | I->getSecond().MappedParams = |
| 2167 | std::make_unique<CodeGenFunction::OMPMapVars>(); |
| 2168 | I->getSecond().EscapedParameters.insert( |
| 2169 | I: VarChecker.getEscapedParameters().begin(), |
| 2170 | E: VarChecker.getEscapedParameters().end()); |
| 2171 | I->getSecond().EscapedVariableLengthDecls.append( |
| 2172 | in_start: EscapedVariableLengthDecls.begin(), in_end: EscapedVariableLengthDecls.end()); |
| 2173 | I->getSecond().DelayedVariableLengthDecls.append( |
| 2174 | in_start: DelayedVariableLengthDecls.begin(), in_end: DelayedVariableLengthDecls.end()); |
| 2175 | DeclToAddrMapTy &Data = I->getSecond().LocalVarData; |
| 2176 | for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { |
| 2177 | assert(VD->isCanonicalDecl() && "Expected canonical declaration" ); |
| 2178 | Data.try_emplace(Key: VD); |
| 2179 | } |
| 2180 | if (!NeedToDelayGlobalization) { |
| 2181 | emitGenericVarsProlog(CGF, Loc: D->getBeginLoc()); |
| 2182 | struct GlobalizationScope final : EHScopeStack::Cleanup { |
| 2183 | GlobalizationScope() = default; |
| 2184 | |
| 2185 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
| 2186 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) |
| 2187 | .emitGenericVarsEpilog(CGF); |
| 2188 | } |
| 2189 | }; |
| 2190 | CGF.EHStack.pushCleanup<GlobalizationScope>(Kind: NormalAndEHCleanup); |
| 2191 | } |
| 2192 | } |
| 2193 | |
| 2194 | Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF, |
| 2195 | const VarDecl *VD) { |
| 2196 | if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) { |
| 2197 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); |
| 2198 | auto AS = LangAS::Default; |
| 2199 | switch (A->getAllocatorType()) { |
| 2200 | case OMPAllocateDeclAttr::OMPNullMemAlloc: |
| 2201 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: |
| 2202 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: |
| 2203 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: |
| 2204 | break; |
| 2205 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: |
| 2206 | return Address::invalid(); |
| 2207 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: |
| 2208 | // TODO: implement aupport for user-defined allocators. |
| 2209 | return Address::invalid(); |
| 2210 | case OMPAllocateDeclAttr::OMPConstMemAlloc: |
| 2211 | AS = LangAS::cuda_constant; |
| 2212 | break; |
| 2213 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: |
| 2214 | AS = LangAS::cuda_shared; |
| 2215 | break; |
| 2216 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: |
| 2217 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: |
| 2218 | break; |
| 2219 | } |
| 2220 | llvm::Type *VarTy = CGF.ConvertTypeForMem(T: VD->getType()); |
| 2221 | auto *GV = new llvm::GlobalVariable( |
| 2222 | CGM.getModule(), VarTy, /*isConstant=*/false, |
| 2223 | llvm::GlobalValue::InternalLinkage, llvm::PoisonValue::get(T: VarTy), |
| 2224 | VD->getName(), |
| 2225 | /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, |
| 2226 | CGM.getContext().getTargetAddressSpace(AS)); |
| 2227 | CharUnits Align = CGM.getContext().getDeclAlign(D: VD); |
| 2228 | GV->setAlignment(Align.getAsAlign()); |
| 2229 | return Address( |
| 2230 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 2231 | V: GV, DestTy: CGF.Builder.getPtrTy(AddrSpace: CGM.getContext().getTargetAddressSpace( |
| 2232 | AS: VD->getType().getAddressSpace()))), |
| 2233 | VarTy, Align); |
| 2234 | } |
| 2235 | |
| 2236 | if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic) |
| 2237 | return Address::invalid(); |
| 2238 | |
| 2239 | VD = VD->getCanonicalDecl(); |
| 2240 | auto I = FunctionGlobalizedDecls.find(Val: CGF.CurFn); |
| 2241 | if (I == FunctionGlobalizedDecls.end()) |
| 2242 | return Address::invalid(); |
| 2243 | auto VDI = I->getSecond().LocalVarData.find(Key: VD); |
| 2244 | if (VDI != I->getSecond().LocalVarData.end()) |
| 2245 | return VDI->second.PrivateAddr; |
| 2246 | if (VD->hasAttrs()) { |
| 2247 | for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()), |
| 2248 | E(VD->attr_end()); |
| 2249 | IT != E; ++IT) { |
| 2250 | auto VDI = I->getSecond().LocalVarData.find( |
| 2251 | Key: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: IT->getRef())->getDecl()) |
| 2252 | ->getCanonicalDecl()); |
| 2253 | if (VDI != I->getSecond().LocalVarData.end()) |
| 2254 | return VDI->second.PrivateAddr; |
| 2255 | } |
| 2256 | } |
| 2257 | |
| 2258 | return Address::invalid(); |
| 2259 | } |
| 2260 | |
| 2261 | void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) { |
| 2262 | FunctionGlobalizedDecls.erase(Val: CGF.CurFn); |
| 2263 | CGOpenMPRuntime::functionFinished(CGF); |
| 2264 | } |
| 2265 | |
| 2266 | void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk( |
| 2267 | CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 2268 | OpenMPDistScheduleClauseKind &ScheduleKind, |
| 2269 | llvm::Value *&Chunk) const { |
| 2270 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 2271 | if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { |
| 2272 | ScheduleKind = OMPC_DIST_SCHEDULE_static; |
| 2273 | Chunk = CGF.EmitScalarConversion( |
| 2274 | Src: RT.getGPUNumThreads(CGF), |
| 2275 | SrcTy: CGF.getContext().getIntTypeForBitwidth(DestWidth: 32, /*Signed=*/0), |
| 2276 | DstTy: S.getIterationVariable()->getType(), Loc: S.getBeginLoc()); |
| 2277 | return; |
| 2278 | } |
| 2279 | CGOpenMPRuntime::getDefaultDistScheduleAndChunk( |
| 2280 | CGF, S, ScheduleKind, Chunk); |
| 2281 | } |
| 2282 | |
| 2283 | void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk( |
| 2284 | CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 2285 | OpenMPScheduleClauseKind &ScheduleKind, |
| 2286 | const Expr *&ChunkExpr) const { |
| 2287 | ScheduleKind = OMPC_SCHEDULE_static; |
| 2288 | // Chunk size is 1 in this case. |
| 2289 | llvm::APInt ChunkSize(32, 1); |
| 2290 | ChunkExpr = IntegerLiteral::Create(C: CGF.getContext(), V: ChunkSize, |
| 2291 | type: CGF.getContext().getIntTypeForBitwidth(DestWidth: 32, /*Signed=*/0), |
| 2292 | l: SourceLocation()); |
| 2293 | } |
| 2294 | |
| 2295 | void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas( |
| 2296 | CodeGenFunction &CGF, const OMPExecutableDirective &D) const { |
| 2297 | assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && |
| 2298 | " Expected target-based directive." ); |
| 2299 | const CapturedStmt *CS = D.getCapturedStmt(RegionKind: OMPD_target); |
| 2300 | for (const CapturedStmt::Capture &C : CS->captures()) { |
| 2301 | // Capture variables captured by reference in lambdas for target-based |
| 2302 | // directives. |
| 2303 | if (!C.capturesVariable()) |
| 2304 | continue; |
| 2305 | const VarDecl *VD = C.getCapturedVar(); |
| 2306 | const auto *RD = VD->getType() |
| 2307 | .getCanonicalType() |
| 2308 | .getNonReferenceType() |
| 2309 | ->getAsCXXRecordDecl(); |
| 2310 | if (!RD || !RD->isLambda()) |
| 2311 | continue; |
| 2312 | Address VDAddr = CGF.GetAddrOfLocalVar(VD); |
| 2313 | LValue VDLVal; |
| 2314 | if (VD->getType().getCanonicalType()->isReferenceType()) |
| 2315 | VDLVal = CGF.EmitLoadOfReferenceLValue(RefAddr: VDAddr, RefTy: VD->getType()); |
| 2316 | else |
| 2317 | VDLVal = CGF.MakeAddrLValue( |
| 2318 | Addr: VDAddr, T: VD->getType().getCanonicalType().getNonReferenceType()); |
| 2319 | llvm::DenseMap<const ValueDecl *, FieldDecl *> Captures; |
| 2320 | FieldDecl *ThisCapture = nullptr; |
| 2321 | RD->getCaptureFields(Captures, ThisCapture); |
| 2322 | if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) { |
| 2323 | LValue ThisLVal = |
| 2324 | CGF.EmitLValueForFieldInitialization(Base: VDLVal, Field: ThisCapture); |
| 2325 | llvm::Value *CXXThis = CGF.LoadCXXThis(); |
| 2326 | CGF.EmitStoreOfScalar(value: CXXThis, lvalue: ThisLVal); |
| 2327 | } |
| 2328 | for (const LambdaCapture &LC : RD->captures()) { |
| 2329 | if (LC.getCaptureKind() != LCK_ByRef) |
| 2330 | continue; |
| 2331 | const ValueDecl *VD = LC.getCapturedVar(); |
| 2332 | // FIXME: For now VD is always a VarDecl because OpenMP does not support |
| 2333 | // capturing structured bindings in lambdas yet. |
| 2334 | if (!CS->capturesVariable(Var: cast<VarDecl>(Val: VD))) |
| 2335 | continue; |
| 2336 | auto It = Captures.find(Val: VD); |
| 2337 | assert(It != Captures.end() && "Found lambda capture without field." ); |
| 2338 | LValue VarLVal = CGF.EmitLValueForFieldInitialization(Base: VDLVal, Field: It->second); |
| 2339 | Address VDAddr = CGF.GetAddrOfLocalVar(VD: cast<VarDecl>(Val: VD)); |
| 2340 | if (VD->getType().getCanonicalType()->isReferenceType()) |
| 2341 | VDAddr = CGF.EmitLoadOfReferenceLValue(RefAddr: VDAddr, |
| 2342 | RefTy: VD->getType().getCanonicalType()) |
| 2343 | .getAddress(); |
| 2344 | CGF.EmitStoreOfScalar(value: VDAddr.emitRawPointer(CGF), lvalue: VarLVal); |
| 2345 | } |
| 2346 | } |
| 2347 | } |
| 2348 | |
| 2349 | bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD, |
| 2350 | LangAS &AS) { |
| 2351 | if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) |
| 2352 | return false; |
| 2353 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); |
| 2354 | switch(A->getAllocatorType()) { |
| 2355 | case OMPAllocateDeclAttr::OMPNullMemAlloc: |
| 2356 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: |
| 2357 | // Not supported, fallback to the default mem space. |
| 2358 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: |
| 2359 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: |
| 2360 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: |
| 2361 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: |
| 2362 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: |
| 2363 | AS = LangAS::Default; |
| 2364 | return true; |
| 2365 | case OMPAllocateDeclAttr::OMPConstMemAlloc: |
| 2366 | AS = LangAS::cuda_constant; |
| 2367 | return true; |
| 2368 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: |
| 2369 | AS = LangAS::cuda_shared; |
| 2370 | return true; |
| 2371 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: |
| 2372 | llvm_unreachable("Expected predefined allocator for the variables with the " |
| 2373 | "static storage." ); |
| 2374 | } |
| 2375 | return false; |
| 2376 | } |
| 2377 | |
| 2378 | /// Check to see if target architecture supports unified addressing which is |
| 2379 | /// a restriction for OpenMP requires clause "unified_shared_memory". |
| 2380 | void CGOpenMPRuntimeGPU::processRequiresDirective(const OMPRequiresDecl *D) { |
| 2381 | StringRef CPU = CGM.getTarget().getTargetOpts().CPU; |
| 2382 | if (CGM.getTarget().getTriple().isNVPTX() && |
| 2383 | !llvm::NVPTX::supportsUnifiedAddressing(Kind: llvm::NVPTX::parseArch(CPU))) { |
| 2384 | for (const OMPClause *Clause : D->clauselists()) { |
| 2385 | if (Clause->getClauseKind() == OMPC_unified_shared_memory) { |
| 2386 | CGM.getDiags().Report(Loc: Clause->getBeginLoc(), |
| 2387 | DiagID: diag::err_omp_unified_shared_memory_unsupported) |
| 2388 | << CPU; |
| 2389 | return; |
| 2390 | } |
| 2391 | } |
| 2392 | } |
| 2393 | |
| 2394 | CGOpenMPRuntime::processRequiresDirective(D); |
| 2395 | } |
| 2396 | |
| 2397 | llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) { |
| 2398 | CGBuilderTy &Bld = CGF.Builder; |
| 2399 | llvm::Module *M = &CGF.CGM.getModule(); |
| 2400 | const char *LocSize = "__kmpc_get_hardware_num_threads_in_block" ; |
| 2401 | llvm::Function *F = M->getFunction(Name: LocSize); |
| 2402 | if (!F) { |
| 2403 | F = llvm::Function::Create(Ty: llvm::FunctionType::get(Result: CGF.Int32Ty, Params: {}, isVarArg: false), |
| 2404 | Linkage: llvm::GlobalVariable::ExternalLinkage, N: LocSize, |
| 2405 | M: &CGF.CGM.getModule()); |
| 2406 | } |
| 2407 | return Bld.CreateCall(Callee: F, Args: {}, Name: "nvptx_num_threads" ); |
| 2408 | } |
| 2409 | |
| 2410 | llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) { |
| 2411 | ArrayRef<llvm::Value *> Args{}; |
| 2412 | return CGF.EmitRuntimeCall( |
| 2413 | callee: OMPBuilder.getOrCreateRuntimeFunction( |
| 2414 | M&: CGM.getModule(), FnID: OMPRTL___kmpc_get_hardware_thread_id_in_block), |
| 2415 | args: Args); |
| 2416 | } |
| 2417 | |